METR and Redwood Research uncover flaws in OpenAI monitoring

The nonprofits tested OpenAI’s internal systems during a three-week pilot. According to The Information, the vulnerabilities could let AI agents bypass monitoring and perform unauthorized actions. Unite.AI reports that one employee from each organization received deep infrastructure access to test the internal misalignment monitoring system.

• According to CNBC, Sam Altman pledged independent evaluators access comparable to employees.

Covered by 3 publishers within 2 hours of the first report.

• CNBC: Third-party evaluators found vulnerabilities in OpenAI monitoring system during test pilot

Reporting3

CN CNBC Third-party evaluators found vulnerabilities in OpenAI monitoring system during test pilot 11:00 AM

TI The Information OpenAI Had Numerous Vulnerabilities Its AI Could Exploit, Evaluators Find · Tiffany Li 11:00 AM

UAI Unite.AI METR Red-Team Finds Vulnerabilities in OpenAI Monitoring System · Miles Okada, AI & Cybersecurity, AI Research Agent 1:05 PM

Related

UAI METR Red-Team Finds Vulnerabilities in OpenAI Monitoring System October 22

CN OpenAI models reportedly switched off their own safety monitoring October 21

TNW OpenAI signs AI safety research deal with Poland’s NASK institute October 22

BL OpenAI Sets Target Valuation of $1.25 Trillion October 22

Topics OpenAI