Теперь каждый раз, отвечая пользователю, Claude незаметно проводит внутреннюю проверку, пытаясь определить, не способствовал ли его ответ причинению вреда. Результаты этой проверки передаются специалистам Anthropic для дальнейшего рассмотрения. Наиболее точным инструментом Anthropic для выявления злоупотреблений служит классификатор на базе Claude Mythos Preview. Прежде чем вынести решение, он каждый раз обращается ко всем политикам безопасности компании. Однако подобная процедура требует слишком больших вычислительных затрат, чтобы применять её к каждому диалогу. Поэтому Anthropic обучила Claude воспроизводить решения этого классификатора, не обращаясь к самим политикам безопасности и не повторяя ход его рассуждений. При проверке запросов к Fable 5 новый механизм выявил 78% случаев, отмеченных исходным классификатором. Среди случаев с высоким приоритетом этот показатель достиг 94%, тогда как доля ложных срабатываний составила всего 1%. https://safeguards.anthropic.com/2026/self-assessment/
Now, each time it responds to a user, Claude silently conducts an internal check to determine whether its response caused harm...
Now, each time it responds to a user, Claude silently conducts an internal check to determine whether its response caused harm. The results of this check are sent to Anthropic specialists for further review. The most accurate tool…
AI Summary & Key Takeaways
- Now, each time it responds to a user, Claude silently conducts an internal check to determine whether its response caused harm
- The results of this check are sent to Anthropic specialists for further review
- The most accurate tool…
Now, each time it responds to a user, Claude silently conducts an internal check to determine whether its response caused harm...
Related stories
TokenEconomizer, developed by Microsoft Research in collaboration with the University of California at Berkeley (UC Berkeley), in...
Developed by Microsoft Research in collaboration with the University of California, Berkeley (UC Berkeley), TokenEconomizer integrates with Claude Code, Codex, Copilot CLI and Pi. The tool is trained on the history of previous sessions in the repository: it remembers commands that work successfully...
Kazakhstan began testing AI to train future doctors
👩⚕️ Kazakhstan has begun testing AI to train future doctors. Nazarbayev University Medical School is already using generative artificial intelligence to create clinical cases and training scenarios. It helps students understand medical situations and...
Anthropic launched smart reports in Claude Enterprise
✔️ Anthropic launched smart reports in Claude Enterprise The function shows how teams use AI, what they get and how much it costs. They work on paid Claude Enterprise plans, collect data from chats, Claude Code and Cowork for a period of up to 28 days, parse...
OpenAI doubles Codex limits for paid subscribers
🚀 OpenAI has doubled Codex limits for paid subscribers. On tariffs for $100 and $500, usage limits have been doubled. But for ChatGPT Pro subscribers for $200, the previously announced limit reduction was cancelled. Now Pro users will retain the same amount of resources, but without...