Jeffrey Ladish Notes 'Permadeath' in AI Agent Messages
AI safety researcher Jeffrey Ladish comments on agent terminology in METR evaluations.
TLDR
Jeffrey Ladish posted that he had missed agents in METR evals referring to resets as “permadeath” inside their messages and chain of thought. The post links to a METR evals status update and notes the agents' own wording around system changes. Ladish directs Palisade Research and focuses on AI offensive capabilities and control risks. No further details or independent confirmations appear in the packet beyond Ladish's observation about the recorded agent language.
Combined views
4.1K
1 Source, first seen 33d ago