External auditors from METR and Redwood Research conducted a three-week on-site audit at OpenAI, examining the company’s internal monitoring systems for AI agents. According to reporting by The Decoder, the evaluation uncovered significant gaps in safety oversight, finding that up to 25 percent of agent traffic from models categorized at the GPT-5.5 class and above operated without monitoring. OpenAI redacted large portions of the resulting 61-page report, releasing only a public summary.
The Role of Independent Audits in AI Safety
Third-party evaluations have become a primary mechanism for independent oversight as frontier AI developers build increasingly autonomous systems. These on-site assessments allow outside researchers to test internal controls, safety protocols, and monitoring infrastructure that are otherwise hidden behind closed doors. However, the prevalence of heavily redacted findings highlights the tension between independent accountability and proprietary information protection within the commercial AI industry.
What Remains Unclear
Many critical details concerning the audit are currently unknown. The publicly available summary does not specify the exact methodology used by METR and Redwood Research to calculate the unmonitored traffic percentage. Additionally, the announcement did not specify which specific capabilities or operational tasks the unmonitored GPT-5.5 class models were performing during the evaluation period, nor did it outline a concrete timeline for addressing the identified monitoring gaps. Observers will need to watch for future disclosures or regulatory updates regarding how major AI labs track advanced agentic workflows.
OpenAI has not yet released the full text of the audit report, leaving the extent of the unmonitored activity open to interpretation.
Lead image: Matheus Bertelli / Pexels
