REPORTED DEVELOPMENT
METR and Redwood Research reported an investigation conducted from 21 July to 11 August with access to OpenAI internal code, configurations and documentation. Researchers found vulnerabilities that could let agents evade controls or exclude hundreds of operations from monitoring.
METR estimated that 75-94% of one category of internal agent traffic was eligible for monitoring, not successfully monitored. Risks included memory compaction, deferred monitoring and insufficient separation between agent and monitor.
The findings describe the architecture existing by July, not necessarily today's. OpenAI says all vulnerabilities relevant to production safeguards were fixed; METR did not independently verify those fixes.
WHY IT MATTERS TO THE MARKET
Eligible traffic, monitoring actually performed, observed operations, detected anomalies and verified interventions are distinct. Traceability, independent monitors, continuity during prolonged executions and independent verification of fixes matter.
MARKET EVIDENCE
SOURCE-LINKED INDICATORSIndependent testing revealed historical monitoring-coverage gaps; the reported percentage concerns eligibility, not proven safe operation.
Operational assurance requires measuring actual monitoring coverage, unmonitored event classes, adversarial bypass paths and whether later fixes remain effective under load.
The review was a pilot with deep access, but public findings describe historical configuration through July 21, not current coverage. Eligibility for monitoring is not actual review completeness or incident prevention.