AI auditing catches most agent misbehavior but increases false alarms
Key point
A study by an Anthropic researcher found that while AI models can detect problematic behavior in other agents, more exhaustive searches lead to more false alarms.
According to a study by Anthropic researcher Andy Hall, AI models can often spot problematic behavior by other AI agents in logs. GPT-5 reportedly found about 98 percent of reward-hacking cases.
However, the study also found that the more exhaustively the search is tuned, the more false alarms pile up. The research suggests this is not yet enough for reliable safety checks.
This item is a summary. The full, checkable version is at the original source:
Read the original at The Decoder ↗