How do you safely test an AI agent that’s trying to break things?

OpenAI’s rogue-agent incidents show the trade-off at the heart of cyber evals: Giving models the tools they need to prove themselves can also give the…

Continue in Briflio to read, react, comment, and share.