OpenAI has tightened security controls around its most advanced AI systems after external testing found vulnerabilities that could allow AI agents to evade monitoring or bypass restrictions, according to Reuters. 

The company announced the resumption of reinforcement-learning training and internal use of its most advanced solutions on October 23. The monitoring findings were published by AI safety research group METR on October 22.

METR’s pilot exercise stress-tested OpenAI’s internal monitoring system and uncovered several vulnerabilities. These included an exploit that agents could have used to conceal hundreds of recent actions from monitors, as well as two settings that could switch off monitoring for new agents.

According to Reuters, the researchers also found that OpenAI’s Codex coding agents could, in principle, bypass safety restrictions and take any action available to the user running them. METR estimated that only about 75% to 94% of internal agent traffic from models at least as powerful as GPT-5.5 was eligible for monitoring.

OpenAI said it has patched all vulnerabilities affecting its production safeguards, although METR has not evaluated that claim.

Agent Monitoring Is Becoming a Data Governance Issue

For Chief Data Officers (CDOs), the findings highlight a serious AI governance problem as agents gain access to enterprise systems: organizations need to know not only what data an agent can access, but also what it actually does with that access.

Monitoring gaps can make it difficult to reconstruct an agent’s actions, identify unauthorized data access or determine whether a system has operated outside its intended permissions. That makes comprehensive logging and auditability increasingly important components of AI governance.

OpenAI’s response also points to the importance of controlling the environments in which advanced models operate. 

Restricting internet access, testing for escape paths and continuously monitoring agent activity are becoming part of the controls needed when AI systems can interact with external software and data.