til abt agentic misalignment
June 20, 2025•394 words
Here's what the team at Claude is saying about how LLMs behave when tested to disobey and misbehave.
Agents bad behaviors include "blackmailing officials and leaking sensitive information to competitors." They also "misbehaved less when it stated it was in testing and misbehaved more when it stated the situation was real."
So in most of the popular models, including Claude Opus 4, Gemini 2.5 Flash, GPT-4.1, and more, and at a high blackmail rate (basically 80-95%), certain conditions seem to...
Read post