1 source filed · EN 1
Set-up: 30 API operations with an append-only audit log, 19 scored tasks plus two unauthorised-refund traps, baseline on Claude Sonnet 5.5 over three runs; thresholds set at twice baseline variance.
1 source filed · EN 1
Pooled prediction across tasks improves modestly with early execution prefixes, because it mostly picks up task difficulty.
0.50mean within-task AUC of a fine-tuned monitor at 20k tokens over 32,406 attempts, i.e. chance
1 source filed · EN 1
Roughly 75-94% of agentic internal Responses API traffic from GPT-5.5-class or stronger models was eligible for monitoring; 99.2% of internal Codex agent traffic was.
roughly 75-94%share of agentic internal Responses API traffic (GPT-5.5-class and stronger) eligible for monitoring at the time of the exercise
1 source filed · EN 1
Argues logs, traces and telemetry need security-grade protection: recent misalignment incidents were caught only through evidence in reasoning traces and logs. Relevant to agent-platform logging design.