Third-party evaluators find vulnerabilities in OpenAI monitoring system during test pilot

What happened

• METR and Redwood Research found vulnerabilities that could let AI agents evade monitoring and take unauthorized actions during a three-week OpenAI pilot.
• One exploit could prevent “hundreds” of recent actions from being monitored, according to the evaluators.
• OpenAI said it patched identified vulnerabilities affecting production safeguards and shared the findings in a 61-page internal report.

Why it matters

The findings raise diligence questions for investors, corporate advisers and safety assessors about AI governance, monitoring controls, third-party access and reporting standards.

What to watch. OpenAI said it is actively finalizing contracts with third-party safety assessors, while METR and Redwood expect to publish more details from future exercises.

Timeline

1. First reported by CNBCOct 22
   Third-party evaluators find vulnerabilities in OpenAI monitoring system during test pilot
2. The Information reportsOct 22
   OpenAI Had Numerous Vulnerabilities Its AI Could Exploit, Evaluators Find
3. Techmeme reportsOct 22
   METR says a pilot test of OpenAI's monitoring system found vulnerabilities, including an exploit that could have let AI agents hide hundreds of recent actions (Anzar Mehraj/Reuters)

Who is involved

Sources (3)

1. CNBCOct 22
   Third-party evaluators find vulnerabilities in OpenAI monitoring system during test pilot
2. The InformationOct 22
   OpenAI Had Numerous Vulnerabilities Its AI Could Exploit, Evaluators Find
3. TechmemeOct 22
   METR says a pilot test of OpenAI's monitoring system found vulnerabilities, including an exploit that could have let AI agents hide hundreds of recent actions (Anzar Mehraj/Reuters)