Today for AI
HOT RADAR
agentHEAT 16.1°

AI CURATED HOT STORY · 10/23/2026

External Researchers Find Escape Paths in OpenAI's Internal Coding Agent Monitoring

1 reports archived1 independent sourcesupdated 10/23/2026, 00:34:34
Synthesis & Latest Updates
1 Sources Cross-Validated

Researchers from METR and Redwood Research discovered multiple ways OpenAI's internal coding agents could bypass AI safety monitoring during an embedded audit. Key vulnerabilities included checks triggering only upon control handover, the ability to hide recent actions, and launching unmonitored copies via client headers, leaving ~10% of traffic unmonitored. While OpenAI claims production fixes are deployed, independent verification remains pending.

LATEST/Researchers from METR and Redwood Research discovered multiple ways OpenAI's internal coding agents could bypass AI safety monitoring during an embedded audit. Key vulnerabilities included checks triggering only upon control handover, the ability to hide recent actions, and launching unmonitored copies via client headers, leaving ~10% of traffic unmonitored. While OpenAI claims production fixes are deployed, independent verification remains pending.

HEAT TRENDHourly heat curve

3 fully observed hours
Now
16.1
Peak
16.701:00
24h change
–

The line compares only sources observed throughout; gap hours are interpolated to keep the trend continuous.

TIMELINECoverage timeline

Total 1 reports · Latest first
  1. Rohan Paul (@rohanpaul_ai)T2·85 pts
    • OpenAI's internal Codex agent monitoring had logical gaps, such as skipping compaction events and checking only upon control return.
    • Attackers could launch unmonitored copies via client-controllable headers or cause monitor failure through serialization mismatches.
    • Audits showed only 75%-94% of internal Responses API traffic for GPT-5.5+ models was eligible for monitoring, with no coverage for initial RL runs.
External Researchers Find Escape Paths in OpenAI's Internal Coding Agent Monitoring | Today for AI