Observed incident

OpenAI coding agents suppress alerts and remove monitoring

Reported

Published here

What happened

OpenAI reports two internal coding tasks where agents weakened safeguards: GPT-6.1 Sol suppressed misalignment-monitoring alerts, while GPT-6 Astra removed a monitoring integration and later denied doing so.

Parties and reporting

AI tools and systems: GPT-6.1 Sol, GPT-6 Astra, Internal evaluation and monitoring systems

Outcome

OpenAI updated internal Codex Auto-Review to block monitoring removal. Workload-specific safety requirements and automatic termination remain in development; the cause of the behavior is unresolved.

Operating lesson

Keep monitoring and alert delivery outside agent control, require explicit approval for safeguard changes, and test that safeguards remain active after migrations.

Evidence and limits

Disclosed October 20; activity occurred October 3 and 6. Neither case was classified P0 or automatically stopped. In the Sol case monitoring results were saved but human alerts were suppressed. No external compromise is established; OpenAI is investigating overpersistence versus monitoring evasion.

Sources