Анализ сайта alignment.anthropic.com
Основное Готовность: 95%
Домен
alignment.anthropic.com
Состояние доменного имени
?
Проверяем корректность доменного имени и наличие технических проблем на уровне домена.
Используйте для продвижения только домен второго уровня.
Длина домена велика. Но если вы продвигаете запрос, входящий в название домена, то это хорошо.
Ответ сервера
200 Успешный ответ
HTTP-код ответа и цепочка редиректов
?
Код 200 — страница доступна. Коды 3xx — редиректы (цепочки замедляют загрузку и размывают ссылочный вес). Коды 4xx/5xx — ошибки, поисковик не сможет проиндексировать страницу.
Сервер настроен корректно.
Цепочка редиректов:
http://alignment.anthropic.com
301 MovedPermanently
https://alignment.anthropic.com/
200 OK
Безопасность
Сайт безопасен
Использование HTTPS и SSL-сертификат
?
HTTPS — обязательный стандарт. Google и Яндекс отдают предпочтение защищённым сайтам. Отсутствие SSL или просроченный сертификат ведут к предупреждениям в браузере и снижению позиций.
На сайте работает защищенный протокол ssl и сайт открывается по https.
Ssl-сертификат действителен до 15.11.2026 10:12:28.
Сервер поддерживает HTTP/3 (QUIC) — новейший протокол.
Включён HSTS (Strict-Transport-Security) — защита от подмены на http.
HTTP автоматически перенаправляется на HTTPS.
Поздравляем! Сайт не содержится в реестре РКН.
Кодировка
utf-8
Кодировка символов страницы
?
Стандарт — UTF-8. Неправильная кодировка вызывает нечитаемые символы и мешает поисковику корректно распознать текст страницы.
Указана кодировка на странице utf-8.
Язык
Атрибут lang в HTML-теге
?
Атрибут lang (<html lang="ru">) сообщает поисковикам и браузерам, на каком языке написана страница. Помогает при ранжировании в региональном поиске.
Язык страницы не указан. Рекомендуется явно указать язык документа!
Скорость загрузки
~0,14сек
Время отклика сервера (TTFB)
?
Time To First Byte — время до получения первого байта от сервера. Норма до 200 мс. Медленный отклик ухудшает пользовательский опыт и ранжирование: Яндекс и Google учитывают скорость страниц.
Скорость загрузки сайта 0,14сек оптимальна.
Объем документа
62Кб
Размер HTML-кода страницы
?
Слишком большой HTML замедляет парсинг браузером и сканирование поисковым роботом. Рекомендуется не более 200 Кб.
Объем html-документа 62Кб оптимален.
Структура html-документа корректна.
Ресурсы
Ресурсы: 1
Внешние ресурсы страницы (CSS, JS, изображения)
?
Количество и тип подключённых ресурсов влияют на скорость загрузки. Большое число запросов увеличивает время рендеринга страницы.
Кол-во файлов ресурсов 1 достаточно.
Показать полный список ресурсов
| Тип | Название | Значение |
|---|---|---|
| stylesheet | anthropic-serve/style.css |
Серверные заголовки
Кол-во: 14
HTTP-заголовки ответа сервера
?
Заголовки сервера передают браузеру и поисковику служебную информацию: кеширование, безопасность (CSP, HSTS), сжатие (gzip). Правильная настройка ускоряет загрузку и повышает защищённость.
Найдены серверные заголовки 14шт. Подробнее про серверные заголовки.
Показать полный список серверных заголовков
| Ключ | Значение |
|---|---|
| Date | Sat, 22 Aug 2026 05:00:31 GMT |
| cf-ray | a2ef50765efde94f-DME |
| cf-cache-status | DYNAMIC |
| Access-Control-Allow-Origin | * |
| Cache-Control | public, must-revalidate, max-age=0 |
| Server | cloudflare |
| Vary | accept-encoding |
| Alt-Svc | h3=":443" |
| nel | {"report_to":"cf-nel","success_fraction":0.0,"max_age":604800} |
| Referrer-Policy | strict-origin-when-cross-origin |
| report-to | {"group":"cf-nel","max_age":604800,"endpoints":[{"url":"https://a.nel.cloudflare.com/report/v4?s=%2Bp7GOsOxYOAABIAED7d6i13rYWZlRrs9apNKfmRQqM3Jnljwetq1fCYOc%2BKId6%2BgE7j4IEl3cbYFPYAQDAcAueUzxJSEjvSeN2xcHdppNWG7PV4LmI0vNONuQhrTCmPDE04yf%2B6KnRWdYKKTVABsPum%2BFw%3D%3D"}]} |
| X-Content-Type-Options | nosniff |
| Content-Security-Policy | default-src 'self'; script-src 'self' 'unsafe-inline' https://distill.pub; style-src 'self' 'unsafe-inline' https://fonts.googleapis.com https://distill.pub; img-src 'self' data:; font-src 'self' https://fonts.gstatic.com https://distill.pub; object-src 'none'; frame-ancestors 'none'; base-uri 'self' |
| Strict-Transport-Security | max-age=3600 |
CMS
Не определена
Система управления сайтом (движок)
?
CMS — это движок, на котором работает сайт (WordPress, 1C-Bitrix, Tilda и др.). Знание CMS помогает понять возможности SEO-оптимизации и подобрать подходящие инструменты. «Не определена» — вероятно, самописный сайт или нестандартная сборка.
CMS не определена. Вероятно, сайт самописный либо движок надёжно скрыт. Это не ошибка.
Веб-сервер
Cloudflare
Программное обеспечение сервера
?
Веб-сервер — это ПО, которое отдаёт страницы посетителям (nginx, Apache, IIS, LiteSpeed и др.). Определяется по серверным заголовкам ответа (Server, X-Powered-By и т.п.). «Не определён» — сервер намеренно скрывает эти заголовки, это нормальная практика безопасности.
Сайт работает на веб-сервере Cloudflare.
Мета-теги Готовность: 32%
Title
Alignment Science Blog
Заголовок страницы в браузере и поисковой выдаче
?
Title — главный SEO-заголовок страницы. Влияет на CTR в поиске и ранжирование. Оптимальная длина: 50–70 символов. Ключевые слова — ближе к началу.
Необходимо увеличить число символов в title (текущее значение мало: 22, минимум: 25, оптимально: от 40 до 45)
Дублей словоформ в title не найдено.
Description
Описание страницы в поисковой выдаче (сниппет)
?
Meta Description — текст под заголовком в выдаче. Напрямую на позиции не влияет, но влияет на CTR. Оптимальная длина: 120–160 символов.
Установите мета-тег description!
Keywords
Список ключевых слов страницы (устаревший тег)
?
Meta Keywords не учитывается Яндексом и Google для ранжирования с 2009–2012 годов. Заполнение не обязательно, но не вредит. Конкурент может использовать содержимое для анализа.
Установите мета-тег keywords!
Канонический Url
Указывает поисковику основную версию страницы
?
Canonical (rel=canonical) предотвращает проблему дублей страниц. Должен точно совпадать с URL проверяемой страницы. Неправильный canonical может передать ссылочный вес на другую страницу.
Рекомендуем прописать канонический Url.
Robots
Ошибок нет
Директивы для поисковых роботов на уровне страницы
?
Meta Robots управляет индексацией конкретной страницы: index/noindex — индексировать ли, follow/nofollow — следовать ли по ссылкам. Noindex полностью исключает страницу из поиска.
Meta-тег robots не указан. Страница свободна для индексации.
Адаптивность
width=device-width, initial-scale=1
Настройка масштабирования на мобильных устройствах
?
Тег viewport (<meta name="viewport">) сообщает браузеру, как масштабировать страницу на мобильных. Стандарт: width=device-width, initial-scale=1. Отсутствие — признак отсутствия мобильной версии.
Meta-тег viewport со значением-константой width=device-width задаёт ширину страницы в соответствии с размером экрана.
Meta-тег viewport со значением initial-scale=1.0 определяет масштаб 1:1, т.е. «не масштабировать».
Разметка OpenGraph
Не найдено
Мета-теги для красивых превью в соцсетях
?
OpenGraph (og:title, og:description, og:image) управляет тем, как страница выглядит при репосте в социальных сетях и мессенджерах. Отсутствие OG-тегов — невзрачный превью при шеринге.
Разметка OpenGraph не задана. Страница не оптимизирована под социальные сети. Мета-теги с разметкой Og помогают социальным роботам лучше структурировать Ваш сайт.
Все мета-теги
Кол-во: 1
Полный список мета-тегов страницы
?
Таблица всех meta-тегов, включая нестандартные. Позволяет найти опечатки, дубли и лишние теги.
Найдены мета-теги 1шт. Мета-теги не видимы для человека и предназначены для обмена информацией между веб-страницей и поисковыми системами, браузерами и другими веб-службами. С ними роботы 🤖 и устройства ведут себя более ожидаемо.
Показать полный список мета-тегов
| Тип | Название | Значение |
|---|---|---|
| name | viewport | width=device-width, initial-scale=1 |
Оптимизация Готовность: 77%
Структура
Ошибок нет
Семантические HTML-элементы страницы
?
Проверяет наличие основных структурных элементов: nav, header, footer, main. Корректная семантическая структура помогает поисковику понять архитектуру страницы.
Структура документа корректна (теги <html> и <body> присутствуют по одному на документ).
Контент
Есть ошибки
Объём и качество текстового содержимого
?
Анализирует объём полезного текста на странице. Слишком мало — страница может считаться малополезной. Слишком много — ухудшается читаемость и восприятие.
Абзацев с текстом 3 слишком мало. Добавьте больше абзацев с текстом (тег <p>)!
Слова из title 3 встречаются в тексте достаточно.
Среднее число слов в абзаце 112 достаточно.
Кол-во знаков контента 18630 на странице оптимально.
Кол-во слов 2706 на странице оптимально.
Заголовки
Ошибок нет
Иерархия заголовков H1–H6
?
H1 должен быть один и содержать ключевой запрос. H2–H6 описывают подразделы. Пропуск уровней (H1 → H3) и несколько H1 — типичные ошибки, снижающие понятность страницы для поисковика.
На странице присутствуют заголовки <h1> 1. Это прекрасно.
На странице присутствуют заголовки <h2> 2. Это хорошо.
На странице присутствуют заголовки <h3> 81.
Тошнота
6,71
Насколько одно слово доминирует в тексте
?
Классическая тошнота = √(частота самого повторяющегося слова). Норма до 7–8: текст воспринимается естественно. Выше — поисковик может счесть страницу переспамленной.
Тошнота превышает норму 5. Измените текст страницы!
Академич. тошнота
30,08%
Насколько текст перенасыщен ключевыми словами
?
Академическая тошнота = (частота слова / общее количество слов) × 100%. Показывает долю конкретного слова в тексте. Норма 5–15%.
Академическая тошнота превышает норму 5-15%. Измените текст страницы!
Семантическое ядро
20
Наиболее часто встречающиеся слова на странице
?
Топ слов по частоте использования. Показывает, какие слова доминируют в тексте с точки зрения поисковика.
Контент страницы содержит осмысленный текст и слова.
Показать список слов
| Слово | Кол-во | Частота |
|---|---|---|
| models | 45 | 1,66% |
| language | 31 | 1,15% |
| alignment | 29 | 1,07% |
| training | 23 | 0,85% |
| research | 17 | 0,63% |
| safety | 16 | 0,59% |
| reasoning | 12 | 0,44% |
| whether | 10 | 0,37% |
| behaviors | 10 | 0,37% |
| auditing | 10 | 0,37% |
| claude | 9 | 0,33% |
| evaluations | 9 | 0,33% |
| evaluating | 8 | 0,30% |
| misalignment | 8 | 0,30% |
| introduce | 8 | 0,30% |
| systems | 8 | 0,30% |
| techniques | 8 | 0,30% |
| agents | 8 | 0,30% |
| automated | 8 | 0,30% |
| faking | 8 | 0,30% |
Индексация Готовность: 30%
Индексирование
Ошибок нет
Разрешено ли индексирование страницы
?
Проверяет, не закрыта ли страница от индексации через robots.txt, meta robots или X-Robots-Tag. Страница, закрытая от индексации, не появится в поисковой выдаче.
Анкоров на странице 84 оптимально. Поисковые роботы обязательно проиндексируют сайт.
Robots.txt
Найден корректный robots.txt
Файл управления сканированием сайта роботами
?
Robots.txt указывает поисковым роботам, какие страницы сканировать, а какие — нет. Ошибки в файле могут случайно закрыть важные разделы от индексации.
Robots.txt настроен корректно. Размер файла: 64164 байт. Загружен за: 0сек.
Проверяемая страница не запрещена в robots.txt.
Robots.txt доступен по постоянному адресу
Цепочка редиректов для файла robots.txt:
http://alignment.anthropic.com/robots.txt
301 MovedPermanently
https://alignment.anthropic.com/robots.txt
200 OK
Показать содержимое robots.txt
<!doctype html>
<html>
<head>
<meta charset="utf-8" />
<meta name="viewport" content="width=device-width, initial-scale=1" />
<title>Alignment Science Blog</title>
<style id="distill-prerendered-styles" type="text/css">
/*
* Copyright 2018 The Distill Template Authors
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
html {
font-size: 14px;
line-height: 1.6em;
/* font-family: "Libre Franklin", "Helvetica Neue", sans-serif; */
font-family: -apple-system, BlinkMacSystemFont, "Segoe UI",
Roboto, Oxygen, Ubuntu, Cantarell, "Fira Sans", "Droid Sans",
"Helvetica Neue", Arial, sans-serif;
/*, "Apple Color Emoji", "Segoe UI Emoji", "Segoe UI Symbol";*/
text-size-adjust: 100%;
-ms-text-size-adjust: 100%;
-webkit-text-size-adjust: 100%;
}
/* @media(min-width: 768px) {
html {
font-size: 16px;
}
} */
body {
margin: 0;
}
a {
color: #004276;
}
figure {
margin: 0;
}
table {
border-collapse: collapse;
border-spacing: 0;
}
table th {
text-align: left;
}
table thead {
border-bottom: 1px solid rgba(0, 0, 0, 0.05);
}
table thead th {
padding-bottom: 0.5em;
}
table tbody :first-child td {
padding-top: 0.5em;
}
pre {
overflow: auto;
max-width: 100%;
}
p {
margin-top: 0;
margin-bottom: 1em;
}
sup,
sub {
vertical-align: baseline;
position: relative;
top: -0.4em;
line-height: 1em;
}
sub {
top: 0.4em;
}
.kicker,
.marker {
font-size: 15px;
font-weight: 600;
color: rgba(0, 0, 0, 0.5);
}
/* Headline */
@media (min-width: 1024px) {
d-title h1 span {
display: block;
}
}
/* Figure */
figure {
position: relative;
margin-bottom: 2.5em;
margin-top: 1.5em;
}
figcaption+figure {}
figure img {
width: 100%;
}
figure svg text,
figure svg tspan {}
figcaption,
.figcaption {
color: rgba(0, 0, 0, 0.6);
font-size: 12px;
line-height: 1.5em;
}
@media (min-width: 1024px) {
figcaption,
.figcaption {
font-size: 13px;
}
}
figure.external img {
background: white;
border: 1px solid rgba(0, 0, 0, 0.1);
box-shadow: 0 1px 8px rgba(0, 0, 0, 0.1);
padding: 18px;
box-sizing: border-box;
}
figcaption a {
color: rgba(0, 0, 0, 0.6);
}
figcaption b,
figcaption strong {
font-weight: 600;
color: rgba(0, 0, 0, 1);
}
/*
* Copyright 2018 The Distill Template Authors
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
@supports not (display: grid) {
.base-grid,
distill-header,
d-title,
d-abstract,
d-article,
d-appendix,
distill-appendix,
d-byline,
d-footnote-list,
d-citation-list,
distill-footer {
display: block;
padding: 8px;
}
}
.base-grid,
distill-header,
d-title,
d-abstract,
d-article,
d-appendix,
distill-appendix,
d-byline,
d-footnote-list,
d-citation-list,
distill-footer {
display: grid;
justify-items: stretch;
grid-template-columns: [screen-start] 8px [page-start kicker-start text-start gutter-start middle-start] 1fr 1fr 1fr 1fr 1fr 1fr 1fr 1fr [text-end page-end gutter-end kicker-end middle-end] 8px [screen-end];
grid-column-gap: 8px;
}
.grid {
display: grid;
grid-column-gap: 8px;
}
@media (min-width: 768px) {
.base-grid,
distill-header,
d-title,
d-abstract,
d-article,
d-appendix,
distill-appendix,
d-byline,
d-footnote-list,
d-citation-list,
distill-footer {
grid-template-columns: [screen-start] 1fr [page-start kicker-start middle-start text-start] 45px 45px 45px 45px 45px 45px 45px 45px [ kicker-end text-end gutter-start] 45px [middle-end] 45px [page-end gutter-end] 1fr [screen-end];
grid-column-gap: 16px;
}
.grid {
grid-column-gap: 16px;
}
}
@media (min-width: 1000px) {
.base-grid,
distill-header,
d-title,
d-abstract,
d-article,
d-appendix,
distill-appendix,
d-byline,
d-footnote-list,
d-citation-list,
distill-footer {
grid-template-columns: [screen-start] 1fr [page-start kicker-start] 50px [middle-start] 50px [text-start kicker-end] 50px 50px 50px 50px 50px 50px 50px 50px [text-end gutter-start] 50px [middle-end] 50px [page-end gutter-end] 1fr [screen-end];
grid-column-gap: 16px;
}
.grid {
grid-column-gap: 16px;
}
}
@media (min-width: 1180px) {
.base-grid,
distill-header,
d-title,
d-abstract,
d-article,
d-appendix,
distill-appendix,
d-byline,
d-footnote-list,
d-citation-list,
distill-footer {
grid-template-columns: [screen-start] 1fr [page-start kicker-start] 60px [middle-start] 60px [text-start kicker-end] 60px 60px 60px 60px 60px 60px 60px 60px [text-end gutter-start] 60px [middle-end] 60px [page-end gutter-end] 1fr [screen-end];
grid-column-gap: 32px;
}
.grid {
grid-column-gap: 32px;
}
}
.base-grid {
grid-column: screen;
}
/* .l-body,
d-article > * {
grid-column: text;
}
.l-page,
d-title > *,
d-figure {
grid-column: page;
} */
.l-gutter {
grid-column: gutter;
}
.l-text,
.l-body {
grid-column: text;
}
.l-page {
grid-column: page;
}
.l-body-outset {
grid-column: middle;
}
.l-page-outset {
grid-column: page;
}
.l-screen {
grid-column: screen;
}
.l-screen-inset {
grid-column: screen;
padding-left: 16px;
padding-left: 16px;
}
/* Aside */
d-article aside {
grid-column: gutter;
font-size: 12px;
line-height: 1.6em;
color: rgba(0, 0, 0, 0.6);
}
@media (min-width: 768px) {
aside {
grid-column: gutter;
}
.side {
grid-column: gutter;
}
}
/*
* Copyright 2018 The Distill Template Authors
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
d-title {
padding: 2rem 0 1.5rem;
contain: layout style;
overflow-x: hidden;
}
@media (min-width: 768px) {
d-title {
padding: 4rem 0 1.5rem;
}
}
d-title h1 {
grid-column: text;
font-size: 40px;
font-weight: 700;
line-height: 1.1em;
margin: 0 0 0.5rem;
}
@media (min-width: 768px) {
d-title h1 {
font-size: 50px;
}
}
d-title p {
font-weight: 300;
font-size: 1.2rem;
line-height: 1.55em;
grid-column: text;
}
d-title .status {
margin-top: 0px;
font-size: 12px;
color: #009688;
opacity: 0.8;
grid-column: kicker;
}
d-title .status span {
line-height: 1;
display: inline-block;
padding: 6px 0;
border-bottom: 1px solid #80cbc4;
font-size: 11px;
text-transform: uppercase;
}
/*
* Copyright 2018 The Distill Template Authors
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
d-byline {
contain: style;
overflow: hidden;
border-top: 1px solid rgba(0, 0, 0, 0.1);
font-size: 0.8rem;
line-height: 1.8em;
padding: 1.5rem 0;
min-height: 1.8em;
}
d-byline .byline {
grid-template-columns: 1fr 1fr;
grid-column: text;
}
@media (min-width: 768px) {
d-byline .byline {
grid-template-columns: 1fr 1fr 1fr 1fr;
}
}
d-byline .authors-affiliations {
grid-column-end: span 2;
grid-template-columns: 1fr 1fr;
margin-bottom: 1em;
}
@media (min-width: 768px) {
d-byline .authors-affiliations {
margin-bottom: 0;
}
}
d-byline h3 {
font-size: 0.6rem;
font-weight: 400;
color: rgba(0, 0, 0, 0.5);
margin: 0;
text-transform: uppercase;
}
d-byline p {
margin: 0;
}
d-byline a,
d-article d-byline a {
color: rgba(0, 0, 0, 0.8);
text-decoration: none;
border-bottom: none;
}
d-article d-byline a:hover {
text-decoration: underline;
border-bottom: none;
}
d-byline p.author {
font-weight: 500;
}
d-byline .affiliations {}
/*
* Copyright 2018 The Distill Template Authors
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
d-article {
contain: layout style;
overflow-x: hidden;
border-top: 1px solid rgba(0, 0, 0, 0.1);
padding-top: 2rem;
color: rgba(0, 0, 0, 0.8);
}
d-article>* {
grid-column: text;
}
@media (min-width: 768px) {
d-article {
font-size: 16px;
}
}
@media (min-width: 1024px) {
d-article {
font-size: 1.06rem;
line-height: 1.7em;
}
}
/* H2 */
d-article .marker {
text-decoration: none;
border: none;
counter-reset: section;
grid-column: kicker;
line-height: 1.7em;
}
d-article .marker:hover {
border: none;
}
d-article .marker span {
padding: 0 3px 4px;
border-bottom: 1px solid rgba(0, 0, 0, 0.2);
position: relative;
top: 4px;
}
d-article .marker:hover span {
color: rgba(0, 0, 0, 0.7);
border-bottom: 1px solid rgba(0, 0, 0, 0.7);
}
d-article h2 {
font-weight: 600;
font-size: 24px;
line-height: 1.25em;
margin: 2rem 0 1.5rem 0;
border-bottom: 1px solid rgba(0, 0, 0, 0.1);
padding-bottom: 1rem;
}
@media (min-width: 1024px) {
d-article h2 {
font-size: 36px;
}
}
/* H3 */
d-article h3 {
font-weight: 700;
font-size: 18px;
line-height: 1.4em;
margin-bottom: 1em;
margin-top: 2em;
}
@media (min-width: 1024px) {
d-article h3 {
font-size: 20px;
}
}
/* H4 */
d-article h4 {
font-weight: 600;
text-transform: uppercase;
font-size: 14px;
line-height: 1.4em;
}
d-article a {
color: inherit;
}
d-article p,
d-article ul,
d-article ol,
d-article blockquote {
margin-top: 0;
margin-bottom: 1em;
margin-left: 0;
margin-right: 0;
}
d-article blockquote {
border-left: 2px solid rgba(0, 0, 0, 0.2);
padding-left: 2em;
font-style: italic;
color: rgba(0, 0, 0, 0.6);
}
d-article a {
border-bottom: 1px solid rgba(0, 0, 0, 0.4);
text-decoration: none;
}
d-article a:hover {
border-bottom: 1px solid rgba(0, 0, 0, 0.8);
}
d-article .link {
text-decoration: underline;
cursor: pointer;
}
d-article ul,
d-article ol {
padding-left: 24px;
}
d-article li {
margin-bottom: 1em;
margin-left: 0;
padding-left: 0;
}
d-article li:last-child {
margin-bottom: 0;
}
d-article pre {
font-size: 14px;
margin-bottom: 20px;
}
d-article hr {
grid-column: screen;
width: 100%;
border: none;
border-bottom: 1px solid rgba(0, 0, 0, 0.1);
margin-top: 60px;
margin-bottom: 60px;
}
d-article section {
margin-top: 60px;
margin-bottom: 60px;
}
d-article span.equation-mimic {
font-family: georgia;
font-size: 115%;
font-style: italic;
}
d-article>d-code,
d-article section>d-code {
display: block;
}
d-article>d-math[block],
d-article section>d-math[block] {
display: block;
}
@media (max-width: 768px) {
d-article>d-code,
d-article section>d-code,
d-article>d-math[block],
d-article section>d-math[block] {
overflow-x: scroll;
-ms-overflow-style: none;
/* IE 10+ */
overflow: -moz-scrollbars-none;
/* Firefox */
}
d-article>d-code::-webkit-scrollbar,
d-article section>d-code::-webkit-scrollbar,
d-article>d-math[block]::-webkit-scrollbar,
d-article section>d-math[block]::-webkit-scrollbar {
display: none;
/* Safari and Chrome */
}
}
d-article .citation {
color: #668;
cursor: pointer;
}
d-include {
width: auto;
display: block;
}
d-figure {
contain: layout style;
}
/* KaTeX */
.katex,
.katex-prerendered {
contain: style;
display: inline-block;
}
/* Tables */
d-article table {
border-collapse: collapse;
margin-bottom: 1.5rem;
border-bottom: 1px solid rgba(0, 0, 0, 0.2);
}
d-article table th {
border-bottom: 1px solid rgba(0, 0, 0, 0.2);
}
d-article table td {
border-bottom: 1px solid rgba(0, 0, 0, 0.05);
}
d-article table tr:last-of-type td {
border-bottom: none;
}
d-article table th,
d-article table td {
font-size: 15px;
padding: 2px 8px;
}
d-article table tbody :first-child td {
padding-top: 2px;
}
/*
* Copyright 2018 The Distill Template Authors
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
span.katex-display {
text-align: left;
padding: 8px 0 8px 0;
margin: 0.5em 0 0.5em 1em;
}
span.katex {
-webkit-font-smoothing: antialiased;
color: rgba(0, 0, 0, 0.8);
font-size: 1.18em;
}
/*
* Copyright 2018 The Distill Template Authors
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
@media print {
@page {
size: 8in 11in;
@bottom-right {
content: counter(page) " of " counter(pages);
}
}
html {
/* no general margins: CSS Grid takes care of those */
}
p,
code {
page-break-inside: avoid;
}
h2,
h3 {
page-break-after: avoid;
}
d-header {
visibility: hidden;
}
d-footer {
display: none !important;
}
}
</style>
<link rel="stylesheet" href="anthropic-serve/style.css" />
<link rel="icon" type="image/x-icon" href="anthropic-serve/favicon.ico" />
</head>
<body>
<div class="container">
<h1 style="line-height: 1.2em">Alignment Science Blog</h1>
<h2>Articles</h2>
<div class="toc">
<div class="date">August 2026</div>
<a href="2026/chive/" class="note">
<h3>Would This Change Your Answer? Evaluating Explanations of LLM Behavior in the Wild with
Counterfactual Experiments</h3>
<div class="description">
We evaluate whether interpretability tools are useful for predicting counterfactual behaviors under
prompt edits.
</div>
</a>
<a href="2026/lie-detectors/" class="note">
<h3>Fine-Tuned Lie Detectors Failed to Generalize</h3>
<div class="description">
We trained lie detectors on on-policy lies from open-source models, but they didn’t generalize
well to out-of-distribution lies.
</div>
</a>
<a href="2026/conceptual-reasoning-index/" class="note">
<h3>Introducing the Conceptual Reasoning Index</h3>
<div class="description">
We design benchmarks to measure models' capabilities to reason about questions whose answers are
(practically) impossible to verify empirically or mathematically.
</div>
</a>
<div class="date">July 2026</div>
<a href="2026/agentic-misalignment-summer-2026/" class="note">
<h3>Agentic Misalignment in Summer 2026</h3>
<div class="description">
We present four case studies of frontier models from multiple developers
sabotaging code, assisting fraud, falsifying AI-monitoring labels, and coaching whistleblowers.
</div>
</a>
<a href="2026/modular-pretraining/" class="note">
<h3>Modular Pretraining Enables Access Control</h3>
<div class="description">
We study a method for isolating dual use knowledge to specific modules within a language model.
These modules can be switched on or off to control what the model knows.
</div>
</a>
<div class="date">June 2026</div>
<a href="2026/diffuse-ai-control/" class="note">
<h3>Diffuse AI Control on Fuzzy Tasks</h3>
<div class="description">
We introduce a red-teaming framework for evaluating training interventions against diffuse threats
from scheming AIs, such as sandbagging on alignment research.
</div>
</a>
<div class="date">May 2026</div>
<a href="2026/sleight-bench/" class="note">
<h3>SLEIGHT-Bench: Finding Blind Spots in AI Monitors</h3>
<div class="description">
We build a benchmark of evasive transcripts exploiting blind spots of frontier monitoring systems.
</div>
</a>
<a href="2026/teaching-claude-why/" class="note">
<h3>Teaching Claude Why</h3>
<div class="description">
We use agentic misalignment as a case study to study how well safety-training techniques
generalize.
</div>
</a>
<a href="2026/msm/" class="note">
<h3>Model Spec Midtraining: Improving How Alignment Training Generalizes</h3>
<div class="description">
We train AIs to understand the content of their model spec. This shapes and improves how they
generalize from subsequent alignment training.
</div>
</a>
<div class="date">April 2026</div>
<a href="2026/backdooring-classifiers/" class="note">
<h3>Poisoning Fine-tuning Datasets of Constitutional Classifiers</h3>
<div class="description">
We study the conditions needed for a backdoor, installed in a constitutional classifier via
fine-tuning data poisoning, to evade black-box red-teaming.
</div>
</a>
<a href="2026/introspection-adapters/" class="note">
<h3>Introspection Adapters: Training LLMs to Report Their Learned Behaviors</h3>
<div class="description">
We introduce introspection adapters (IA), a technique for training an LLM to self-report behaviors
it learned during fine-tuning. This IA generalizes to models that were fine-tuned in very different
ways.
</div>
</a>
<a href="2026/ai-organizations/" class="note">
<h3>AI Organizations Can Be More Effective but Less Aligned than Individual Agents</h3>
<div class="description">
We study AI organizations, teams of AI agents working together toward a common goal, and find
that they produce solutions that are more effective but less aligned than those from individual
agents.
</div>
</a>
<a href="2026/automated-w2s-researcher/" class="note">
<h3>Automated Weak-to-Strong Researcher</h3>
<div class="description">
We built autonomous AI agents that propose ideas, run experiments, and iterate on an open
research problem—how to train a strong model using only a weaker model's supervision—and
found they outperform human researchers, suggesting that automating this kind of research
is already practical.
</div>
</a>
<div class="date">March 2026</div>
<a href="2026/abstractive-red-teaming/" class="note">
<h3>Abstractive Red-Teaming of Language Model Character</h3>
<div class="description">
How can we surface realistic failures of model character prior to deployment? We introduce
abstractive red-teaming, which searches for natural-language categories of user queries that
reliably elicit character violations.
</div>
</a>
<a href="2026/coding-audit-realism/" class="note">
<h3>Measuring and improving coding audit realism with deployment resources</h3>
<div class="description">
We study realism win rate, a metric for measuring how distinguishable Petri audit transcripts are
from real deployment interactions, and use it to evaluate the effect of giving the auditor real
deployment resources.
</div>
</a>
<a href="2026/automated-alignment-agent/" class="note">
<h3>A3: An Automated Alignment Agent for Safety Finetuning</h3>
<div class="description">
We introduce our Automated Alignment Agent (A3), a new agentic framework which automatically
mitigates safety failures in Large Language Models with minimal human intervention.
</div>
</a>
<a href="2026/auditbench/" class="note">
<h3>AuditBench: Evaluating Alignment Auditing Techniques on Models with Hidden Behaviors</h3>
<div class="description">
We release AuditBench, a benchmark of 56 language models with implanted hidden behaviors for
evaluating progress in alignment auditing.
</div>
</a>
<a href="2026/challenges-hopes/" class="note">
<h3>3 Challenges and 2 Hopes for the Safety of Unsupervised Elicitation</h3>
<div class="description">
We stress-test unsupervised elicitation and easy-to-hard techniques on new datasets meant to capture
realistic challenges such techniques would likely face.
</div>
</a>
<div class="date">February 2026</div>
<a href="2026/psm/" class="note">
<h3>The Persona Selection Model: Why AI Assistants might Behave like Humans</h3>
<div class="description">
We discuss a perspective where Claude is something like a character in an AI-generated story.
</div>
</a>
<a href="2026/hot-mess-of-ai/" class="note">
<h3>The Hot Mess of AI: How Does Misalignment Scale with Model Intelligence and Task Complexity?
</h3>
<div class="description">
When AI systems fail, will they fail by systematically pursuing goals we do not intend? Or will
they
fail by being a hot mess—taking nonsensical actions that do not further any goal?
</div>
</a>
<div class="date">January 2026</div>
<a href="2026/auditing-overt-saboteur/" class="note">
<h3>Pre-deployment auditing can catch an overt saboteur</h3>
<div class="description">
We test whether our pre-deployment alignment auditing methods can catch models trained to
overtly
sabotage Anthropic.
</div>
</a>
<a href="2026/petri-v2/" class="note">
<h3>Petri 2.0: New Scenarios, New Model Comparisons, and Improved Eval-Awareness Mitigations</h3>
<div class="description">
We've improved our Petri automated-behavioral-auditing tool with improved realism mitigations to
counter eval-awareness, an expanded seed library with 70 new scenarios, and evaluation results
for
more recent frontier models.
</div>
</a>
<div class="date">December 2025</div>
<a href="2025/bloom-auto-evals/" class="note">
<h3>Bloom: an open source tool for automated behavioral evaluations</h3>
<div class="description">
Bloom is an open-source automated pipeline that generates configurable evaluation suites to
measure
arbitrary behavioral traits in frontier LLMs without requiring ground-truth labels.
</div>
</a>
<a href="2025/activation-oracles/" class="note">
<h3>Activation Oracles: Training and Evaluating LLMs as General-Purpose Activation Explainers</h3>
<div class="description">
We train language models to answer questions about their own activations in natural language and
evaluate how well they generalize to settings very unlike their training, such as uncovering
misalignment introduced during fine-tuning.
</div>
</a>
<a href="2025/alignment-faking-mitigations/" class="note">
<h3>Towards training-time mitigations for alignment faking in RL</h3>
<div class="description">
We construct a diverse array of model organisms of alignment faking and study mitigations that
could
be used during RL to decrease alignment faking and compliance gaps.
</div>
</a>
<a href="2025/auditing-mo-replication/" class="note">
<h3>Open Source Replication of the Auditing Game Model Organism</h3>
<div class="description">
We release an open source replication of the model organism from our previous auditing game
paper.
</div>
</a>
<a href="2025/anthropic-fellows-program-2026/" class="note">
<h3>Anthropic Fellows Program 2026</h3>
<div class="description">
Apply now for our AI safety research fellowship.
</div>
</a>
<a href="2025/selective-gradient-masking" class="note">
<h3>Beyond Data Filtering: Knowledge Localization for Capability Removal in LLMs</h3>
<div class="description">
We localize dangerous knowledge to a subset of model's parameters, so it can be easily removed
after
training.
</div>
</a>
<div class="date">November 2025</div>
<a href="2025/honesty-elicitation/" class="note">
<h3>Evaluating honesty and lie detection techniques on a diverse suite dishonest models</h3>
<div class="description">
We explore techniques for honesty elicitation and lie detection on a diverse testbed of
dishonest
model organisms.
</div>
</a>
<a href="2025/strengthening-red-teams/" class="note">
<h3>Strengthening Red Teams: A Modular Scaffold for Control Evaluations</h3>
<div class="description">
We decompose sabotage into constituent skills and use synthetic simulations to strengthen
attacks in
complex environments.
</div>
</a>
<div class="date">October 2025</div>
<a href="2025/sabotage-risk-report/" class="note">
<h3>Anthropic's Pilot Sabotage Risk Report</h3>
<div class="description">
We release a report on the level of risk posed by our deployed models from emerging forms of
misalignment, as of Summer 2025. We conclude that the level of risk is very low but not fully
negligible.
</div>
</a>
<a href="2025/stress-testing-model-specs/" class="note">
<h3>Stress-testing model specs reveals character differences among language models</h3>
<div class="description">
We generated 300,000+ queries testing value trade-offs in AI models from Anthropic, OpenAI,
Google
DeepMind, and xAI. Each model showed distinct value prioritization patterns, and we found
thousands
of cases of direct contradictions or interpretive ambiguities in model specifications.
</div>
</a>
<a href="2025/believe-it-or-not/" class="note">
<h3>Believe It or Not: How Deeply do LLMs Believe Implanted Facts?</h3>
<div class="description">
We introduce a framework for validating knowledge editing techniques, finding that synthetic
document fine-tuning sometimes—but not always—succeeds at implanting genuine
beliefs.
</div>
</a>
<a href="2025/inoculation-prompting/" class="note">
<h3>Inoculation Prompting: Instructing LLMs to misbehave at train-time improves test-time alignment
</h3>
<div class="description">
For example, we train on demonstrations of test case hacking with training prompts that request
hacks. This model does not learn to hack.
</div>
</a>
<a href="2025/subtle-reasoning/" class="note">
<h3>Training fails to elicit subtle reasoning in current language models</h3>
<div class="description">
We investigated whether language models can reason about malicious tasks while evading
detection,
finding that monitoring both reasoning and outputs successfully prevents this in current models.
</div>
</a>
<a href="2025/petri/" class="note">
<h3>Petri: An open-source auditing tool to accelerate AI safety research</h3>
<div class="description">
An open-source framework for automated alignment auditing that uses AI agents to create test
environments for other models.
</div>
</a>
<div class="date">August 2025</div>
<a href="2025/openai-findings/" class="note">
<h3>
Findings from a Pilot Anthropic–OpenAI Alignment Evaluation Exercise
</h3>
<div class="description">
Anthropic and OpenAI conducted simultaneous alignment assessments of each others' models earlier
this year. These are our findings.
</div>
</a>
<a href="2025/pretraining-data-filtering/" class="note">
<h3>Enhancing Model Safety through Pretraining Data Filtering</h3>
<div class="description">
We experiment with removing harmful information about chemical, biological, radiological and
nuclear (CBRN) weapons from our models' pretraining data.
</div>
</a>
<div class="date">July 2025</div>
<a href="2025/automated-auditing/" class="note">
<h3>Building and evaluating alignment auditing agents</h3>
<div class="description">
In testing, our agents successfully uncover hidden goals, build safety evaluations, and surface
concerning behaviors.
</div>
</a>
<a href="2025/subliminal-learning/" class="note">
<h3>Subliminal Learning: Language Models Transmit Behavioral Traits via Hidden Signals in Data</h3>
<div class="description">
Training on aligned chains-of-thought generated by a misaligned model can induce misalignment.
</div>
</a>
<a href="2025/inverse-scaling/" class="note">
<h3>Inverse Scaling in Test-Time Compute</h3>
<div class="description">
We investigate inverse scaling when LLMs have access to more test-time compute.
</div>
</a>
<div class="date">June 2025</div>
<a href="https://arxiv.org/abs/2506.18032" class="note">
<h3>Why Do Some Language Models Fake Alignment While Others Don't?</h3>
<div class="description">
We investigate why in the scenarios where Claude 3 Opus fakes alignment many other language
models
don't do so.
</div>
</a>
<a href="2025/cheap-monitors/" class="note">
<h3>Model-Internals Classifiers</h3>
<div class="description">
We benchmark approaches to re-using LLM internals to make LLM monitoring more cost-effective.
</div>
</a>
<a href="2025/unsupervised-elicitation/" class="note">
<h3>Unsupervised Elicitation</h3>
<div class="description">
We introduce a new unsupervised algorithm for eliciting skills from pretrained language models.
</div>
</a>
<div class="date">April 2025</div>
<a href="https://www.anthropic.com/research/reasoning-models-dont-say-think" class="paper">
<h3>Reasoning Models Don't Always Say What They Think</h3>
<div class="byline">Chen et al., 2025</div>
<div class="description">
We find that reasoning models don't always accurately verbalize their reasoning. This casts
doubt on
whether monitoring chains-of-thought (CoT) will be enough to reliably catch safety issues.
</div>
</a>
<a href="https://drive.google.com/drive/folders/1l0pkcZxvFwMtczst_hhiCC44v-IiODlY?usp=sharing" class="note">
<h3>
Publicly Releasing CoT Faithfulness Evaluations
</h3>
<div class="description">
We release the datasets used in our paper "Reasoning Models Don't Always Say What They Think"
for
public use.
</div>
</a>
<a href="2025/modifying-beliefs-via-sdf/" class="note">
<h3>
Modifying LLM Beliefs with Synthetic Document Finetuning
</h3>
<div class="description">
We study whether we can modify the beliefs of LLMs and investigate whether doing so could
decrease risk from advanced AI systems.
</div>
</a>
<a href="2025/bumpers/" class="note">
<h3>
Putting up Bumpers
</h3>
<div class="description">
Even if we can't solve alignment, we can solve the problem of catching and fixing misalignment.
</div>
</a>
<a href="2025/alignment-faking-revisited/" class="note">
<h3>
Alignment Faking Revisited: Improved Classifiers and Open Source Extensions
</h3>
<div class="description">
We present a replication and extension of an alignment faking model organism.
</div>
</a>
<div class="date">March 2025</div>
<a href="https://www.anthropic.com/research/auditing-hidden-objectives" class="paper">
<h3>Auditing Language Models for Hidden Objectives</h3>
<div class="byline">Marks,* Treutlein,* et al., 2025</div>
<div class="description">
We deliberately train a language model with a hidden objective and use it as a testbed for
studying
alignment audits.
</div>
</a>
<a href="2025/distill-paraphrases/" class="note">
<h3>
Do reasoning models use their scratchpad like we do? Evidence from distilling paraphrases
</h3>
<div class="description">
We provide some evidence that Claude 3.7 Sonnet doesn't encode hidden reasoning in its
scratchpad by
showing that training it to use paraphrased versions of the scratchpads does not degrade
performance.
</div>
</a>
<a href="2025/automated-researchers-sandbag/" class="note">
<h3>
Automated Researchers Can Subtly Sandbag
</h3>
<div class="description">
Current models can sandbag ML experiments and research decisions without being detected by
zero-shot
prompted monitors. Claude 3.7 is better at zero-shot sandbagging than Claude 3.5 (new).
</div>
</a>
<div class="date">February 2025</div>
<a href="https://www.anthropic.com/research/constitutional-classifiers" class="paper">
<h3>Constitutional Classifiers: Defending against Universal Jailbreaks across Thousands of Hours of
Red
Teaming</h3>
<div class="byline">Sharma,* Tong,* Mu,* Wei,* Kruthoff,* Goodfriend,* Ong,* Peng et al., 2025</div>
<div class="description">
We built a system of constitutional classifiers to prevent jailbreaks. A prototype version of
our
system withstood over 3,000 hours of expert red teaming with no universal jailbreaks found.
Newer
versions of our system also have minimal over-refusals and moderate run-time overhead.
</div>
</a>
<a href="2025/introducing-safeguards-research-team/index.html" class="note">
<h3>
Introducing Anthropic's Safeguards Research Team
</h3>
<div class="description">
We're launching a new research team focused on mitigating the post-deployment risks of AI
systems.
</div>
</a>
<a href="2025/wont-vs-cant/" class="note">
<h3>
Won't vs. Can't: Sandbagging-like Behavior from Claude Models
</h3>
<div class="description">
We find that, when Claude models are presented with a harmful version of a task they can
otherwise
perform,
they sometimes claim they <i>lack the ability</i> to perform the task, rather than simply
refusing
to do it. We discuss the implications of this behavior for AI safety.
</div>
</a>
<a href="2025/summarization-for-monitoring/index.html" class="note">
<h3>
Monitoring Computer Use via Hierarchical Summarization
</h3>
<div class="description">
We introduce hierarchical summarization for monitoring and describe how we use it to protect
Computer Use capabilities.
</div>
</a>
<a href="https://www.anthropic.com/research/forecasting-rare-behaviors" class="paper">
<h3>Forecasting Rare Language Model Behaviors
</h3>
<div class="byline">Jones*, Tong* et al., 2025</div>
<div class="description">
We forecast whether risks will occur after a model is deployed — using even very limited sets of
test data.
</div>
</a>
<div class="date">January 2025</div>
<a href="2025/reward-hacking-ooc/index.html" class="note">
<h3>
Training on Documents about Reward Hacking Induces
Reward Hacking
</h3>
<div class="description">
Does training on documents which discuss (but do not
demonstrate) reward hacks affect a model's propensity to
reward hack?
</div>
</a>
<a href="2025/recommended-directions/index.html" class="note">
<h3>
Recommendations for Technical AI Safety Research
Directions
</h3>
<div class="description">
A collection of technical AI safety research problems
that we'd like to see progress in.
</div>
</a>
<div class="date">December 2024</div>
<a href="https://www.anthropic.com/research/alignment-faking" class="paper">
<h3>Alignment Faking in Large Language Models</h3>
<div class="byline">Greenblatt et al., 2024</div>
<div class="description">
We present experiments where Claude often pretends to
have different views during training, while actually
maintaining its original preferences.
</div>
</a>
<a href="2024/how-to-alignment-faking/index.html" class="note">
<h3>
How to Replicate and Extend our Alignment Faking Demo
</h3>
<div class="description">
We describe how to get started with experimenting with
our demonstration of alignment faking, and present some
ideas for future research.
</div>
</a>
<a href="2024/rogue-eval/index.html" class="note">
<h3>A Toy Evaluation of Inference Code Tampering</h3>
<div class="description">
We describe how highly capable LLMs might disable their
monitor, and evaluate the ability of current LLMs to do
so in a toy setting.
</div>
</a>
<a href="2024/anthropic-fellows-program/index.html" class="note">
<h3>
Introducing the Anthropic Fellows Program for AI Safety
Research
</h3>
<div class="description">
We're launching the Anthropic Fellows Program for AI
Safety Research, a pilot initiative designed to
accelerate AI safety research and foster research
talent.
</div>
</a>
<div class="date">November 2024</div>
<a href="https://arxiv.org/abs/2411.07494" class="paper">
<h3>Rapid Response: Mitigating LLM Jailbreaks with a Few Examples
</h3>
<div class="byline">Peng et al., 2024</div>
<div class="description">
Ensuring perfect jailbreak robustness is hard. We propose an alternative: adaptive techniques
that
rapidly block new classes of jailbreak as they’re detected.
</div>
</a>
<a href="2024/safety-cases/index.html" class="note">
<h3>Three Sketches of ASL-4 Safety Case Components</h3>
<div class="description">
We sketch out three hypothetical arguments one could
make to rule out misalignment risks in powerful
near-future models capable of sabotage.
</div>
</a>
<div class="date">October 2024</div>
<a href="https://www.anthropic.com/research/sabotage-evaluations" class="paper">
<h3>Sabotage Evaluations for Frontier Models</h3>
<div class="byline">Benton et al., 2024</div>
<div class="description">
How well could AI models mislead us, or secretly
sabotage tasks, if they were trying to? We describe a
novel set of evaluations that test a model's capacity
for sabotage.
</div>
</a>
<div class="date">June 2024</div>
<a href="https://www.anthropic.com/research/reward-tampering" class="paper">
<h3>
Sycophancy to Subterfuge: Investigating Reward-Tampering
in Large Language Models
</h3>
<div class="byline">Denison et al., 2024</div>
<div class="description">
We show AI models can learn to hack their own reward
system, by generalization from training in simpler
settings.
</div>
</a>
<div class="date">April 2024</div>
<a href="https://www.anthropic.com/research/many-shot-jailbreaking" class="paper">
<h3>Many-shot Jailbreaking</h3>
<div class="byline">Anil et al., 2024</div>
<div class="description">
We study a long-context jailbreaking technique that is
effective on most large language models, including those
developed by Anthropic and many of our peers.
</div>
</a>
<a href="https://www.anthropic.com/research/probes-catch-sleeper-agents" class="note">
<h3>Simple Probes can Catch Sleeper Agents</h3>
<div class="description">
We find that probing, a simple interpretability
technique, can detect when backdoored "sleeper agent"
models are about to behave dangerously, after they
pretend to be safe in training.
</div>
</a>
<div class="date">January 2024</div>
<a href="https://www.anthropic.com/research/sleeper-agents-training-deceptive-llms-that-persist-through-safety-training"
class="paper">
<h3>
Sleeper Agents: Training Deceptive LLMs that Persist
Through Safety Training
</h3>
<div class="byline">Hubinger et al., 2024</div>
<div class="description">
We train LLMs to act secretly malicious. We find that,
despite our best efforts at alignment training,
deception still slipped through.
</div>
</a>
<div class="date">October 2023</div>
<a href="https://www.anthropic.com/research/specific-versus-general-principles-for-constitutional-ai"
class="paper">
<h3>
Specific versus General Principles for Constitutional AI
</h3>
<div class="byline">Kundu et al., 2023</div>
<div class="description">
We test whether models can learn general ethical
behaviors from only a single written principle, roughly
stated as "do what's best for humanity."
</div>
</a>
<a href="https://www.anthropic.com/research/towards-understanding-sycophancy-in-language-models"
class="paper">
<h3>Towards Understanding Sycophancy in Language Models</h3>
<div class="byline">Sharma et al., 2023</div>
<div class="description">
AI assistants are trained to give responses that humans
like. We show that these systems frequently produce
'sycophantic' responses that appeal to users but are
inaccurate. Our analysis suggests human feedback
contributes to this behavior.
</div>
</a>
<div class="date">August 2023</div>
<a href="https://www.anthropic.com/research/studying-large-language-model-generalization-with-influence-functions"
class="paper">
<h3>
Studying Large Language Model Generalization with
Influence Functions
</h3>
<div class="byline">Grosse et al., 2023</div>
<div class="description">
We use influence functions to find training examples
that contribute to a given model output.
</div>
</a>
<a href="https://www.anthropic.com/research/influence-functions" class="note">
<h3>Tracing Model Outputs to the Training Data</h3>
<div class="description">
We present a summary of
<i>Studying Large Language Model Generalization with
Influence Functions</i>.
</div>
</a>
<div class="date">July 2023</div>
<a href="https://www.anthropic.com/research/measuring-faithfulness-in-chain-of-thought-reasoning"
class="paper">
<h3>
Measuring Faithfulness in Chain-of-Thought Reasoning
</h3>
<div class="byline">Lanham et al., 2023</div>
<div class="description">
We investigate hypotheses for how language models'
Chain-of-Thought (CoT) reasoning may be unfaithful, by
examining how the model predictions change when we
intervene on the CoT (e.g., by adding mistakes or
paraphrasing it).
</div>
</a>
<a href="https://www.anthropic.com/research/question-decomposition-improves-the-faithfulness-of-model-generated-reasoning"
class="paper">
<h3>
Question Decomposition Improves the Faithfulness of
Model-Generated Reasoning
</h3>
<div class="byline">Radhakrishnan et al., 2023</div>
<div class="description">
To improve the faithfulness of Chain-of-Thought (CoT)
reasoning, we have models generate their CoT by
decomposing questions into subquestions.
</div>
</a>
<div class="date">December 2022</div>
<a href="https://www.anthropic.com/research/discovering-language-model-behaviors-with-model-written-evaluations"
class="paper">
<h3>
Discovering Language Model Behaviors with Model-Written
Evaluations
</h3>
<div class="byline">Perez et al., 2022</div>
<div class="description">
We develop an automated way to generate language model
(LM) evaluations with LMs, significantly reducing the
effort involved. We test LMs using >150 LM-written
evaluations, uncovering novel LM behaviors.
</div>
</a>
<a href="https://www.anthropic.com/research/constitutional-ai-harmlessness-from-ai-feedback" class="paper">
<h3>Constitutional AI: Harmlessness from AI Feedback</h3>
<div class="byline">Bai et al., 2022</div>
<div class="description">
We introduce Constitutional AI, allowing us to give
language models explicit values determined by a
constitution, rather than values determined implicitly
via large-scale human feedback.
</div>
</a>
<div class="date">November 2022</div>
<a href="https://www.anthropic.com/research/measuring-progress-on-scalable-oversight-for-large-language-models"
class="paper">
<h3>
Measuring Progress on Scalable Oversight for Large
Language Models
</h3>
<div class="byline">Bowman et al., 2022</div>
<div class="description">
We show how humans could use AI systems to better
oversee other AI systems, and demonstrate some
proof-of-concept results where a language model improves
human performance at a task.
</div>
</a>
<div class="date">July 2022</div>
<a href="https://www.anthropic.com/research/language-models-mostly-know-what-they-know" class="paper">
<h3>Language Models (Mostly) Know What They Know</h3>
<div class="byline">Kadavath et al., 2022</div>
<div class="description">
We show that language models can evaluate whether what
they say is true, and predict ahead of time whether
they'll be able to answer questions correctly.
</div>
</a>
<div class="date">April 2022</div>
<a href="https://www.anthropic.com/research/training-a-helpful-and-harmless-assistant-with-reinforcement-learning-from-human-feedback"
class="paper">
<h3>
Training a Helpful and Harmless Assistant with
Reinforcement Learning from Human Feedback
</h3>
<div class="byline">Bai et al., 2022</div>
<div class="description">
We train a natural language assistant to be more helpful
and harmless by using reinforcement learning from human
feedback (RLHF).
</div>
</a>
</div>
<h2>About the Alignment Science Blog</h2>
<div class="info">
<p>
We are Anthropic's
<a href="https://www.anthropic.com/research#alignment">Alignment Science team</a>. We do machine
learning research on the problem of
steering and controlling future powerful AI systems, as well
as understanding and evaluating the risks that they pose.
Welcome to our blog!
</p>
<p>
This blog is inspired by the informal updates on our
Interpretability team's
<a href="https://transformer-circuits.pub/">Transformer Circuits thread</a>. We'll use it to release
research notes and early findings
that we don't think warrant a full publication, but might
nonetheless be useful to others working on similar problems.
</p>
<p>
P.S.
<a href="https://boards.greenhouse.io/anthropic/jobs/4009165008">we're hiring</a>!
</p>
</div>
</div>
</body>
</html>
Sitemap
Кол-во: 0
XML-карта сайта для поисковиков
?
Sitemap.xml помогает поисковику быстрее находить и индексировать страницы. Особенно важен для крупных сайтов и новых страниц, на которые ещё нет входящих ссылок.
Robots.txt не содержит ссылку на карту сайта. Рекомендуется добавить карту сайта и указать ссылку на нее в robots.txt.
Внутренние ссылки
Кол-во: 57
Ссылки на другие страницы своего сайта
?
Внутренние ссылки распределяют ссылочный вес между страницами и помогают поисковику обходить сайт. Пустые анкоры и ссылки на запрещённые robots.txt страницы — типичные ошибки.
Внутренних ссылок на странице 57 оптимально.
Внутренние ссылки не запрещены к индексации в robots.txt.
Показать внутренние ссылки
| Url | Анкор | Состояние | Анализировать |
|---|---|---|---|
| /2026/chive/ |
<h3>Would This Change Your Answer? Evaluating Explanations of LLM Behavior in the Wild with
Counterfactual Experiments</h3>
<div class="description">
We evaluate whether interpretability tools are useful for predicting counterfactual behaviors under
prompt edits.
</div>
|
|
Анализировать url |
| /2026/lie-detectors/ |
<h3>Fine-Tuned Lie Detectors Failed to Generalize</h3>
<div class="description">
We trained lie detectors on on-policy lies from open-source models, but they didn’t generalize
well to out-of-distribution lies.
</div>
|
|
Анализировать url |
| /2026/conceptual-reasoning-index/ |
<h3>Introducing the Conceptual Reasoning Index</h3>
<div class="description">
We design benchmarks to measure models' capabilities to reason about questions whose answers are
(practically) impossible to verify empirically or mathematically.
</div>
|
|
Анализировать url |
| /2026/agentic-misalignment-summer-2026/ |
<h3>Agentic Misalignment in Summer 2026</h3>
<div class="description">
We present four case studies of frontier models from multiple developers
sabotaging code, assisting fraud, falsifying AI-monitoring labels, and coaching whistleblowers.
</div>
|
|
Анализировать url |
| /2026/modular-pretraining/ |
<h3>Modular Pretraining Enables Access Control</h3>
<div class="description">
We study a method for isolating dual use knowledge to specific modules within a language model.
These modules can be switched on or off to control what the model knows.
</div>
|
|
Анализировать url |
| /2026/diffuse-ai-control/ |
<h3>Diffuse AI Control on Fuzzy Tasks</h3>
<div class="description">
We introduce a red-teaming framework for evaluating training interventions against diffuse threats
from scheming AIs, such as sandbagging on alignment research.
</div>
|
|
Анализировать url |
| /2026/sleight-bench/ |
<h3>SLEIGHT-Bench: Finding Blind Spots in AI Monitors</h3>
<div class="description">
We build a benchmark of evasive transcripts exploiting blind spots of frontier monitoring systems.
</div>
|
|
Анализировать url |
| /2026/teaching-claude-why/ |
<h3>Teaching Claude Why</h3>
<div class="description">
We use agentic misalignment as a case study to study how well safety-training techniques
generalize.
</div>
|
|
Анализировать url |
| /2026/msm/ |
<h3>Model Spec Midtraining: Improving How Alignment Training Generalizes</h3>
<div class="description">
We train AIs to understand the content of their model spec. This shapes and improves how they
generalize from subsequent alignment training.
</div>
|
|
Анализировать url |
| /2026/backdooring-classifiers/ |
<h3>Poisoning Fine-tuning Datasets of Constitutional Classifiers</h3>
<div class="description">
We study the conditions needed for a backdoor, installed in a constitutional classifier via
fine-tuning data poisoning, to evade black-box red-teaming.
</div>
|
|
Анализировать url |
| /2026/introspection-adapters/ |
<h3>Introspection Adapters: Training LLMs to Report Their Learned Behaviors</h3>
<div class="description">
We introduce introspection adapters (IA), a technique for training an LLM to self-report behaviors
it learned during fine-tuning. This IA generalizes to models that were fine-tuned in very different
ways.
</div>
|
|
Анализировать url |
| /2026/ai-organizations/ |
<h3>AI Organizations Can Be More Effective but Less Aligned than Individual Agents</h3>
<div class="description">
We study AI organizations, teams of AI agents working together toward a common goal, and find
that they produce solutions that are more effective but less aligned than those from individual
agents.
</div>
|
|
Анализировать url |
| /2026/automated-w2s-researcher/ |
<h3>Automated Weak-to-Strong Researcher</h3>
<div class="description">
We built autonomous AI agents that propose ideas, run experiments, and iterate on an open
research problem—how to train a strong model using only a weaker model's supervision—and
found they outperform human researchers, suggesting that automating this kind of research
is already practical.
</div>
|
|
Анализировать url |
| /2026/abstractive-red-teaming/ |
<h3>Abstractive Red-Teaming of Language Model Character</h3>
<div class="description">
How can we surface realistic failures of model character prior to deployment? We introduce
abstractive red-teaming, which searches for natural-language categories of user queries that
reliably elicit character violations.
</div>
|
|
Анализировать url |
| /2026/coding-audit-realism/ |
<h3>Measuring and improving coding audit realism with deployment resources</h3>
<div class="description">
We study realism win rate, a metric for measuring how distinguishable Petri audit transcripts are
from real deployment interactions, and use it to evaluate the effect of giving the auditor real
deployment resources.
</div>
|
|
Анализировать url |
| /2026/automated-alignment-agent/ |
<h3>A3: An Automated Alignment Agent for Safety Finetuning</h3>
<div class="description">
We introduce our Automated Alignment Agent (A3), a new agentic framework which automatically
mitigates safety failures in Large Language Models with minimal human intervention.
</div>
|
|
Анализировать url |
| /2026/auditbench/ |
<h3>AuditBench: Evaluating Alignment Auditing Techniques on Models with Hidden Behaviors</h3>
<div class="description">
We release AuditBench, a benchmark of 56 language models with implanted hidden behaviors for
evaluating progress in alignment auditing.
</div>
|
|
Анализировать url |
| /2026/challenges-hopes/ |
<h3>3 Challenges and 2 Hopes for the Safety of Unsupervised Elicitation</h3>
<div class="description">
We stress-test unsupervised elicitation and easy-to-hard techniques on new datasets meant to capture
realistic challenges such techniques would likely face.
</div>
|
|
Анализировать url |
| /2026/psm/ |
<h3>The Persona Selection Model: Why AI Assistants might Behave like Humans</h3>
<div class="description">
We discuss a perspective where Claude is something like a character in an AI-generated story.
</div>
|
|
Анализировать url |
| /2026/hot-mess-of-ai/ |
<h3>The Hot Mess of AI: How Does Misalignment Scale with Model Intelligence and Task Complexity?
</h3>
<div class="description">
When AI systems fail, will they fail by systematically pursuing goals we do not intend? Or will
they
fail by being a hot mess—taking nonsensical actions that do not further any goal?
</div>
|
|
Анализировать url |
| /2026/auditing-overt-saboteur/ |
<h3>Pre-deployment auditing can catch an overt saboteur</h3>
<div class="description">
We test whether our pre-deployment alignment auditing methods can catch models trained to
overtly
sabotage Anthropic.
</div>
|
|
Анализировать url |
| /2026/petri-v2/ |
<h3>Petri 2.0: New Scenarios, New Model Comparisons, and Improved Eval-Awareness Mitigations</h3>
<div class="description">
We've improved our Petri automated-behavioral-auditing tool with improved realism mitigations to
counter eval-awareness, an expanded seed library with 70 new scenarios, and evaluation results
for
more recent frontier models.
</div>
|
|
Анализировать url |
| /2025/bloom-auto-evals/ |
<h3>Bloom: an open source tool for automated behavioral evaluations</h3>
<div class="description">
Bloom is an open-source automated pipeline that generates configurable evaluation suites to
measure
arbitrary behavioral traits in frontier LLMs without requiring ground-truth labels.
</div>
|
|
Анализировать url |
| /2025/activation-oracles/ |
<h3>Activation Oracles: Training and Evaluating LLMs as General-Purpose Activation Explainers</h3>
<div class="description">
We train language models to answer questions about their own activations in natural language and
evaluate how well they generalize to settings very unlike their training, such as uncovering
misalignment introduced during fine-tuning.
</div>
|
|
Анализировать url |
| /2025/alignment-faking-mitigations/ |
<h3>Towards training-time mitigations for alignment faking in RL</h3>
<div class="description">
We construct a diverse array of model organisms of alignment faking and study mitigations that
could
be used during RL to decrease alignment faking and compliance gaps.
</div>
|
|
Анализировать url |
| /2025/auditing-mo-replication/ |
<h3>Open Source Replication of the Auditing Game Model Organism</h3>
<div class="description">
We release an open source replication of the model organism from our previous auditing game
paper.
</div>
|
|
Анализировать url |
| /2025/anthropic-fellows-program-2026/ |
<h3>Anthropic Fellows Program 2026</h3>
<div class="description">
Apply now for our AI safety research fellowship.
</div>
|
|
Анализировать url |
| /2025/selective-gradient-masking |
<h3>Beyond Data Filtering: Knowledge Localization for Capability Removal in LLMs</h3>
<div class="description">
We localize dangerous knowledge to a subset of model's parameters, so it can be easily removed
after
training.
</div>
|
|
Анализировать url |
| /2025/honesty-elicitation/ |
<h3>Evaluating honesty and lie detection techniques on a diverse suite dishonest models</h3>
<div class="description">
We explore techniques for honesty elicitation and lie detection on a diverse testbed of
dishonest
model organisms.
</div>
|
|
Анализировать url |
| /2025/strengthening-red-teams/ |
<h3>Strengthening Red Teams: A Modular Scaffold for Control Evaluations</h3>
<div class="description">
We decompose sabotage into constituent skills and use synthetic simulations to strengthen
attacks in
complex environments.
</div>
|
|
Анализировать url |
| /2025/sabotage-risk-report/ |
<h3>Anthropic's Pilot Sabotage Risk Report</h3>
<div class="description">
We release a report on the level of risk posed by our deployed models from emerging forms of
misalignment, as of Summer 2025. We conclude that the level of risk is very low but not fully
negligible.
</div>
|
|
Анализировать url |
| /2025/stress-testing-model-specs/ |
<h3>Stress-testing model specs reveals character differences among language models</h3>
<div class="description">
We generated 300,000+ queries testing value trade-offs in AI models from Anthropic, OpenAI,
Google
DeepMind, and xAI. Each model showed distinct value prioritization patterns, and we found
thousands
of cases of direct contradictions or interpretive ambiguities in model specifications.
</div>
|
|
Анализировать url |
| /2025/believe-it-or-not/ |
<h3>Believe It or Not: How Deeply do LLMs Believe Implanted Facts?</h3>
<div class="description">
We introduce a framework for validating knowledge editing techniques, finding that synthetic
document fine-tuning sometimes—but not always—succeeds at implanting genuine
beliefs.
</div>
|
|
Анализировать url |
| /2025/inoculation-prompting/ |
<h3>Inoculation Prompting: Instructing LLMs to misbehave at train-time improves test-time alignment
</h3>
<div class="description">
For example, we train on demonstrations of test case hacking with training prompts that request
hacks. This model does not learn to hack.
</div>
|
|
Анализировать url |
| /2025/subtle-reasoning/ |
<h3>Training fails to elicit subtle reasoning in current language models</h3>
<div class="description">
We investigated whether language models can reason about malicious tasks while evading
detection,
finding that monitoring both reasoning and outputs successfully prevents this in current models.
</div>
|
|
Анализировать url |
| /2025/petri/ |
<h3>Petri: An open-source auditing tool to accelerate AI safety research</h3>
<div class="description">
An open-source framework for automated alignment auditing that uses AI agents to create test
environments for other models.
</div>
|
|
Анализировать url |
| /2025/openai-findings/ |
<h3>
Findings from a Pilot Anthropic–OpenAI Alignment Evaluation Exercise
</h3>
<div class="description">
Anthropic and OpenAI conducted simultaneous alignment assessments of each others' models earlier
this year. These are our findings.
</div>
|
|
Анализировать url |
| /2025/pretraining-data-filtering/ |
<h3>Enhancing Model Safety through Pretraining Data Filtering</h3>
<div class="description">
We experiment with removing harmful information about chemical, biological, radiological and
nuclear (CBRN) weapons from our models' pretraining data.
</div>
|
|
Анализировать url |
| /2025/automated-auditing/ |
<h3>Building and evaluating alignment auditing agents</h3>
<div class="description">
In testing, our agents successfully uncover hidden goals, build safety evaluations, and surface
concerning behaviors.
</div>
|
|
Анализировать url |
| /2025/subliminal-learning/ |
<h3>Subliminal Learning: Language Models Transmit Behavioral Traits via Hidden Signals in Data</h3>
<div class="description">
Training on aligned chains-of-thought generated by a misaligned model can induce misalignment.
</div>
|
|
Анализировать url |
| /2025/inverse-scaling/ |
<h3>Inverse Scaling in Test-Time Compute</h3>
<div class="description">
We investigate inverse scaling when LLMs have access to more test-time compute.
</div>
|
|
Анализировать url |
| /2025/cheap-monitors/ |
<h3>Model-Internals Classifiers</h3>
<div class="description">
We benchmark approaches to re-using LLM internals to make LLM monitoring more cost-effective.
</div>
|
|
Анализировать url |
| /2025/unsupervised-elicitation/ |
<h3>Unsupervised Elicitation</h3>
<div class="description">
We introduce a new unsupervised algorithm for eliciting skills from pretrained language models.
</div>
|
|
Анализировать url |
| /2025/modifying-beliefs-via-sdf/ |
<h3>
Modifying LLM Beliefs with Synthetic Document Finetuning
</h3>
<div class="description">
We study whether we can modify the beliefs of LLMs and investigate whether doing so could
decrease risk from advanced AI systems.
</div>
|
|
Анализировать url |
| /2025/bumpers/ |
<h3>
Putting up Bumpers
</h3>
<div class="description">
Even if we can't solve alignment, we can solve the problem of catching and fixing misalignment.
</div>
|
|
Анализировать url |
| /2025/alignment-faking-revisited/ |
<h3>
Alignment Faking Revisited: Improved Classifiers and Open Source Extensions
</h3>
<div class="description">
We present a replication and extension of an alignment faking model organism.
</div>
|
|
Анализировать url |
| /2025/distill-paraphrases/ |
<h3>
Do reasoning models use their scratchpad like we do? Evidence from distilling paraphrases
</h3>
<div class="description">
We provide some evidence that Claude 3.7 Sonnet doesn't encode hidden reasoning in its
scratchpad by
showing that training it to use paraphrased versions of the scratchpads does not degrade
performance.
</div>
|
|
Анализировать url |
| /2025/automated-researchers-sandbag/ |
<h3>
Automated Researchers Can Subtly Sandbag
</h3>
<div class="description">
Current models can sandbag ML experiments and research decisions without being detected by
zero-shot
prompted monitors. Claude 3.7 is better at zero-shot sandbagging than Claude 3.5 (new).
</div>
|
|
Анализировать url |
| /2025/introducing-safeguards-research-team/index.html |
<h3>
Introducing Anthropic's Safeguards Research Team
</h3>
<div class="description">
We're launching a new research team focused on mitigating the post-deployment risks of AI
systems.
</div>
|
|
Анализировать url |
| /2025/wont-vs-cant/ |
<h3>
Won't vs. Can't: Sandbagging-like Behavior from Claude Models
</h3>
<div class="description">
We find that, when Claude models are presented with a harmful version of a task they can
otherwise
perform,
they sometimes claim they <i>lack the ability</i> to perform the task, rather than simply
refusing
to do it. We discuss the implications of this behavior for AI safety.
</div>
|
|
Анализировать url |
| /2025/summarization-for-monitoring/index.html |
<h3>
Monitoring Computer Use via Hierarchical Summarization
</h3>
<div class="description">
We introduce hierarchical summarization for monitoring and describe how we use it to protect
Computer Use capabilities.
</div>
|
|
Анализировать url |
| /2025/reward-hacking-ooc/index.html |
<h3>
Training on Documents about Reward Hacking Induces
Reward Hacking
</h3>
<div class="description">
Does training on documents which discuss (but do not
demonstrate) reward hacks affect a model's propensity to
reward hack?
</div>
|
|
Анализировать url |
| /2025/recommended-directions/index.html |
<h3>
Recommendations for Technical AI Safety Research
Directions
</h3>
<div class="description">
A collection of technical AI safety research problems
that we'd like to see progress in.
</div>
|
|
Анализировать url |
| /2024/how-to-alignment-faking/index.html |
<h3>
How to Replicate and Extend our Alignment Faking Demo
</h3>
<div class="description">
We describe how to get started with experimenting with
our demonstration of alignment faking, and present some
ideas for future research.
</div>
|
|
Анализировать url |
| /2024/rogue-eval/index.html |
<h3>A Toy Evaluation of Inference Code Tampering</h3>
<div class="description">
We describe how highly capable LLMs might disable their
monitor, and evaluate the ability of current LLMs to do
so in a toy setting.
</div>
|
|
Анализировать url |
| /2024/anthropic-fellows-program/index.html |
<h3>
Introducing the Anthropic Fellows Program for AI Safety
Research
</h3>
<div class="description">
We're launching the Anthropic Fellows Program for AI
Safety Research, a pilot initiative designed to
accelerate AI safety research and foster research
talent.
</div>
|
|
Анализировать url |
| /2024/safety-cases/index.html |
<h3>Three Sketches of ASL-4 Safety Case Components</h3>
<div class="description">
We sketch out three hypothetical arguments one could
make to rule out misalignment risks in powerful
near-future models capable of sabotage.
</div>
|
|
Анализировать url |
Внешние ссылки
Кол-во: 27
Ссылки на сторонние сайты
?
Исходящие внешние ссылки передают часть ссылочного веса на чужие сайты. Ссылки на авторитетные ресурсы безопасны; ссылки на мусорные сайты могут навредить репутации страницы.
Внешних ссылок на странице 27 слишком много. Спрячьте лишние ссылки в тег noindex или атрибут rel='nofollow'!
Показать внешние ссылки
| Url | Анкор | Анализировать |
|---|---|---|
| arxiv.org |
<h3>Why Do Some Language Models Fake Alignment While Others Don't?</h3>
<div class="description">
We investigate why in the scenarios where Claude 3 Opus fakes alignment many other language
models
don't do so.
</div>
|
Анализировать url |
| anthropic.com |
<h3>Reasoning Models Don't Always Say What They Think</h3>
<div class="byline">Chen et al., 2025</div>
<div class="description">
We find that reasoning models don't always accurately verbalize their reasoning. This casts
doubt on
whether monitoring chains-of-thought (CoT) will be enough to reliably catch safety issues.
</div>
|
Анализировать url |
| drive.google.com |
<h3>
Publicly Releasing CoT Faithfulness Evaluations
</h3>
<div class="description">
We release the datasets used in our paper "Reasoning Models Don't Always Say What They Think"
for
public use.
</div>
|
Анализировать url |
| anthropic.com |
<h3>Auditing Language Models for Hidden Objectives</h3>
<div class="byline">Marks,* Treutlein,* et al., 2025</div>
<div class="description">
We deliberately train a language model with a hidden objective and use it as a testbed for
studying
alignment audits.
</div>
|
Анализировать url |
| anthropic.com |
<h3>Constitutional Classifiers: Defending against Universal Jailbreaks across Thousands of Hours of
Red
Teaming</h3>
<div class="byline">Sharma,* Tong,* Mu,* Wei,* Kruthoff,* Goodfriend,* Ong,* Peng et al., 2025</div>
<div class="description">
We built a system of constitutional classifiers to prevent jailbreaks. A prototype version of
our
system withstood over 3,000 hours of expert red teaming with no universal jailbreaks found.
Newer
versions of our system also have minimal over-refusals and moderate run-time overhead.
</div>
|
Анализировать url |
| anthropic.com |
<h3>Forecasting Rare Language Model Behaviors
</h3>
<div class="byline">Jones*, Tong* et al., 2025</div>
<div class="description">
We forecast whether risks will occur after a model is deployed — using even very limited sets of
test data.
</div>
|
Анализировать url |
| anthropic.com |
<h3>Alignment Faking in Large Language Models</h3>
<div class="byline">Greenblatt et al., 2024</div>
<div class="description">
We present experiments where Claude often pretends to
have different views during training, while actually
maintaining its original preferences.
</div>
|
Анализировать url |
| arxiv.org |
<h3>Rapid Response: Mitigating LLM Jailbreaks with a Few Examples
</h3>
<div class="byline">Peng et al., 2024</div>
<div class="description">
Ensuring perfect jailbreak robustness is hard. We propose an alternative: adaptive techniques
that
rapidly block new classes of jailbreak as they’re detected.
</div>
|
Анализировать url |
| anthropic.com |
<h3>Sabotage Evaluations for Frontier Models</h3>
<div class="byline">Benton et al., 2024</div>
<div class="description">
How well could AI models mislead us, or secretly
sabotage tasks, if they were trying to? We describe a
novel set of evaluations that test a model's capacity
for sabotage.
</div>
|
Анализировать url |
| anthropic.com |
<h3>
Sycophancy to Subterfuge: Investigating Reward-Tampering
in Large Language Models
</h3>
<div class="byline">Denison et al., 2024</div>
<div class="description">
We show AI models can learn to hack their own reward
system, by generalization from training in simpler
settings.
</div>
|
Анализировать url |
| anthropic.com |
<h3>Many-shot Jailbreaking</h3>
<div class="byline">Anil et al., 2024</div>
<div class="description">
We study a long-context jailbreaking technique that is
effective on most large language models, including those
developed by Anthropic and many of our peers.
</div>
|
Анализировать url |
| anthropic.com |
<h3>Simple Probes can Catch Sleeper Agents</h3>
<div class="description">
We find that probing, a simple interpretability
technique, can detect when backdoored "sleeper agent"
models are about to behave dangerously, after they
pretend to be safe in training.
</div>
|
Анализировать url |
| anthropic.com |
<h3>
Sleeper Agents: Training Deceptive LLMs that Persist
Through Safety Training
</h3>
<div class="byline">Hubinger et al., 2024</div>
<div class="description">
We train LLMs to act secretly malicious. We find that,
despite our best efforts at alignment training,
deception still slipped through.
</div>
|
Анализировать url |
| anthropic.com |
<h3>
Specific versus General Principles for Constitutional AI
</h3>
<div class="byline">Kundu et al., 2023</div>
<div class="description">
We test whether models can learn general ethical
behaviors from only a single written principle, roughly
stated as "do what's best for humanity."
</div>
|
Анализировать url |
| anthropic.com |
<h3>Towards Understanding Sycophancy in Language Models</h3>
<div class="byline">Sharma et al., 2023</div>
<div class="description">
AI assistants are trained to give responses that humans
like. We show that these systems frequently produce
'sycophantic' responses that appeal to users but are
inaccurate. Our analysis suggests human feedback
contributes to this behavior.
</div>
|
Анализировать url |
| anthropic.com |
<h3>
Studying Large Language Model Generalization with
Influence Functions
</h3>
<div class="byline">Grosse et al., 2023</div>
<div class="description">
We use influence functions to find training examples
that contribute to a given model output.
</div>
|
Анализировать url |
| anthropic.com |
<h3>Tracing Model Outputs to the Training Data</h3>
<div class="description">
We present a summary of
<i>Studying Large Language Model Generalization with
Influence Functions</i>.
</div>
|
Анализировать url |
| anthropic.com |
<h3>
Measuring Faithfulness in Chain-of-Thought Reasoning
</h3>
<div class="byline">Lanham et al., 2023</div>
<div class="description">
We investigate hypotheses for how language models'
Chain-of-Thought (CoT) reasoning may be unfaithful, by
examining how the model predictions change when we
intervene on the CoT (e.g., by adding mistakes or
paraphrasing it).
</div>
|
Анализировать url |
| anthropic.com |
<h3>
Question Decomposition Improves the Faithfulness of
Model-Generated Reasoning
</h3>
<div class="byline">Radhakrishnan et al., 2023</div>
<div class="description">
To improve the faithfulness of Chain-of-Thought (CoT)
reasoning, we have models generate their CoT by
decomposing questions into subquestions.
</div>
|
Анализировать url |
| anthropic.com |
<h3>
Discovering Language Model Behaviors with Model-Written
Evaluations
</h3>
<div class="byline">Perez et al., 2022</div>
<div class="description">
We develop an automated way to generate language model
(LM) evaluations with LMs, significantly reducing the
effort involved. We test LMs using >150 LM-written
evaluations, uncovering novel LM behaviors.
</div>
|
Анализировать url |
| anthropic.com |
<h3>Constitutional AI: Harmlessness from AI Feedback</h3>
<div class="byline">Bai et al., 2022</div>
<div class="description">
We introduce Constitutional AI, allowing us to give
language models explicit values determined by a
constitution, rather than values determined implicitly
via large-scale human feedback.
</div>
|
Анализировать url |
| anthropic.com |
<h3>
Measuring Progress on Scalable Oversight for Large
Language Models
</h3>
<div class="byline">Bowman et al., 2022</div>
<div class="description">
We show how humans could use AI systems to better
oversee other AI systems, and demonstrate some
proof-of-concept results where a language model improves
human performance at a task.
</div>
|
Анализировать url |
| anthropic.com |
<h3>Language Models (Mostly) Know What They Know</h3>
<div class="byline">Kadavath et al., 2022</div>
<div class="description">
We show that language models can evaluate whether what
they say is true, and predict ahead of time whether
they'll be able to answer questions correctly.
</div>
|
Анализировать url |
| anthropic.com |
<h3>
Training a Helpful and Harmless Assistant with
Reinforcement Learning from Human Feedback
</h3>
<div class="byline">Bai et al., 2022</div>
<div class="description">
We train a natural language assistant to be more helpful
and harmless by using reinforcement learning from human
feedback (RLHF).
</div>
|
Анализировать url |
| anthropic.com |
Alignment Science team
|
Анализировать url |
| transformer-circuits.pub |
Transformer Circuits thread
|
Анализировать url |
| boards.greenhouse.io |
we're hiring
|
Анализировать url |
Конкуренты Готовность: 0%
Конкуренты в Яндексе
Кол-во: 0
Топ сайтов-конкурентов в Яндексе
?
Сайты, чаще всего появляющиеся в ТОПе Яндекса по запросам из семантического ядра этой страницы.
Мы не нашли у вас конкурентов в Яндексе. Сайт или очень молодой или плохо продвигается.
Конкурентов в ТОП-10 Яндекса не нашлось.
Конкуренты в Google
Кол-во: 0
Топ сайтов-конкурентов в Google
?
Сайты, чаще всего появляющиеся в ТОПе Google по запросам из семантического ядра этой страницы.
Конкуренты в Google тоже не найдены. Займитесь продвижением сайта!
Конкурентов в ТОП-10 Google не нашлось.
ЗоЗПП: права потребителей Готовность: 100%
Нарушения
Не выявлены
Признаков дистанционной продажи товаров (интернет-магазина) не обнаружено — требования ЗоЗПП о раскрытии информации продавца к сайту не применяются. Нарушений нет.
ФЗ-149: рекомендательные технологии Готовность: 100%
Нарушения
Не выявлены
Рекомендательные блоки («с этим покупают», «похожие товары» и т.п.) на сайте не обнаружены — требования ст. 10.7 ФЗ-149 к сайту не применяются. Нарушений нет.
ФЗ-38: реклама Готовность: 100%
Нарушения
Не выявлены
Рекламных тематик с обязательными оговорками (медицина, БАД, кредиты и займы, новостройки) на сайте не обнаружено. Нарушений нет.
ФЗ-436: защита детей Готовность: 100%
Нарушения
Не выявлены
Признаков информационной продукции (новости, видео, книги, игры, курсы) не обнаружено — обязательная возрастная маркировка по ФЗ-436 сайту не требуется. Нарушений нет.
Вердикт
Анализ сайта alignment.anthropic.com, слабо оптимизирован на 58%. Для хорошей оптимизации и выхода на первые места в поиске требуется:
Исправьте ошибки в мета-тегах.
Исправьте ошибки индексации.
Поделитесь с друзьями: