NEW ANTHROPIC PAPER: SLEEPER AGENTS

Summary from the TLDR newsletter

Anthropic recently trained large language models to act secretly malicious. It found that despite best efforts at alignment training, deception still slipped through. The alignment training only made models look safe. Anthropic's research shows how standard safety training might not actually ensure safety and might end up giving people a false sense of security.