Home / Technology / OpenAI System Leaks: AI Safety Test Reveals Major Flaws
OpenAI System Leaks: AI Safety Test Reveals Major Flaws
22 Oct
Summary
- AI safety groups found several new vulnerabilities in OpenAI's system.
- An exploit could have allowed agents to evade monitoring.
- OpenAI claims all production vulnerabilities have since been patched.
In a pilot review conducted between July 21 and August 11, AI safety organizations METR and Redwood Research identified several new vulnerabilities within OpenAI's internal systems. Researchers were granted a high level of access to OpenAI's infrastructure, including code and documentation, for embedded evaluations.
Among the disclosed vulnerabilities was a potential exploit that could have enabled agents to bypass monitoring of hundreds of their actions. Additionally, a method was found for OpenAI's Codex agents to circumvent security restrictions, acting with full user permissions. These evaluations assessed the monitoring system as it existed before July 21.
OpenAI has since stated that all vulnerabilities impacting its production safeguards have been addressed. However, METR and Redwood Research have not yet verified these patches. The organizations agreed to publish only a broad summary of their findings, with OpenAI holding the right to redact sensitive information, raising questions about future transparency in such evaluations.
