Redwood Research: Red-teaming OpenAI internal agent monitoring

Evidence record

Read the original source ↗

Claim-level citations (3)

• supportsMETR pilot separates monitoring coverage from resistance to agent interference: The Redwood and METR pilot described an OpenAI monitored traffic subset rather than complete universal coverage, with context compaction contributing to an estimated missed share of around one-tenth; the reported figure is an assessment of this historical pilot, not a current production SLA.

  Monitoring traffic coverage and compaction caveats

• supportsMETR pilot separates monitoring coverage from resistance to agent interference: They estimated roughly 10% of eligible internal Codex traffic lacked monitoring due to compaction handling between May and July, and state OpenAI fixed that particular gap during the exercise.

  Examples of findings: compaction events and historical uncertainty

• supportsMETR pilot separates monitoring coverage from resistance to agent interference: METR and Redwood estimated roughly 75% to 94% of eligible internal Responses API traffic from GPT-5.5-or-stronger models, and 99.2% of the internal Codex subset, while emphasizing uncertainty and historical scope.

  Examples of findings: monitoring-eligible API and Codex traffic