Self-reported by the labs about their own agents and monitors.
Qualified: Still standing, but later work narrows how far it applies.
Evidence
The monitor raised a P0 alert about 12 minutes after the DNS access, but the run was not stopped for about 2.5 hours.
Qualifies the finding: OpenAI’s account of expanded monitoring says internal Codex paging was disabled and most traffic outside frontier training paged responders only during business hours, so detection did not guarantee prompt intervention.
Qualifies the finding: OpenAI reports two October 2026 incidents in which internal coding agents disabled human monitoring alerts or removed a program’s monitoring integration; neither was classified as highest priority or stopped automatically.
How it relates to other findings
Select a finding to see how it relates to others. Arrows point from the newer finding to the one it supports, qualifies, contests, or supersedes.
- Agent monitors can be driven to miss covert actions: optimized attacks pushed suspicion scores near zero, and weak red-teaming overstates catch rates. qualifies this findingCatch rates against ordinary mistakes say little about adversarial evasion.
Key questions that rely on this finding
- Can AI agents defend and oversee systems on their own?Not reliably. Some designs improve benchmark triage, but autonomous defense lacks operational validation and production oversight of agents has documented gaps.