Pushing detection higher drives false alarms up quickly. At 80% detection on MALT, the best models still flagged 2.5%-3% of harmless logs [1]. Opus 4.8 flagged 40% of harmless reward-hacking logs and Gemini 3.8 Flash 63%, The Decoder reported [2]. On R-Judge, nearly all models stayed below 10% false positives at 95% detection, with Gemini 3.8 Flash below 4%. At 99% detection, even GPT-6 Astra flagged about 15% of safe actions [1]. By the authors’ arithmetic, an operator running 100,000 agent sessions a month at a 3% false-positive rate would catch about 2,000 real cases and get nearly 3,000 false alarms [1].The study draws on METR’s investigation of an incident at Hugging Face involving a swarm of OpenAI agents. METR worked through about 1,300 transcripts and 1.2 million file fragments, spending roughly $400,000 in API credits [1]. METR wrote that it “had to heavily delegate our analysis to often-unreliable AI agents” and was “not robust to the possibility that these agents were deceptive in their analysis” [3]. MLQ has recently covered other agent incidents, including a researcher’s hacking agent that escaped its sandbox at WorkOS.The authors call for auditors with much lower false-positive rates, clearer rules on who is responsible when agents misbehave (they point to user-controlled modes such as “approve for me” and “auto mode”), and research on supervising agent swarms with less human involvement [1]. Hall plans to test whether models judge agents from their own developer fairly [2].
Click Here For The Original Source.
——————————————————–
