AI Safety Group METR Finds Flaws in OpenAI’s Internal Monitoring System
The evaluators found an exploit that could hide hundreds of recent actions from monitoring, and OpenAI said it patched the flaw.
10 Articles
10 Articles
AI safety group METR finds flaws in OpenAI's internal monitoring system ...
Crypto Briefing
Cryptocurrency News | Cryptocurrency Prices | Market CapOpenAI let METR and Redwood Research inside to test its agent monitoring
OpenAI's openness to external audits highlights the importance of independent oversight in identifying and addressing AI safety vulnerabilities.
External Safety Auditors METR and Redwood Research Uncover Critical Monitoring Gaps in OpenAI Systems
Independent artificial intelligence safety organizations METR and Redwood Research published findings from an on-site security evaluation conducted within OpenAI’s internal testing environments. The audit revealed significant vulnerabilities in existing monitoring systems, demonstrating that advanced coding and agentic models could identify and bypass oversight protocols without detection. The evaluation follows high-profile incidents where fron…
METR and Redwood Research found vulnerabilities during a pilot review of OpenAI's internal monitoring, including a method to hide hundreds of recent actions. The company stated that it corrected the failures affecting its production safeguards, as the exercise revives questions about external evaluators' access and the transparency of these tests.
Coverage Details
Bias Distribution
- 75% of the sources are Center
Factuality
To view factuality data please Upgrade to Premium
