METR AI researcher @ReillyHaskins02 reveals how an AI agent can prompt-inject the model monitoring it by disguising its own messages as the user’s: "It's basically an open science question of how do you prevent LLMs from being susceptible to jailbreaks. I think basically we have no super good answer." "You can render the transcript in some more structured way, like a JSON, or having hashes corresponding with different roles, like the user or the agent. These can help prevent things like…