AI NewsWaveExternal Evaluation Finds Monitoring Vulnerabilities in OpenAI Agents
Key points
- METR and Redwood review OpenAI's internal monitoring system
- Vulnerabilities found that allow bypassing oversight and disabling security limits
- OpenAI also built an offline test environment
The AI safety evaluation groups METR and Redwood Research reviewed OpenAI's internal AI agent monitoring system and found multiple vulnerabilities. AI Times reported they are vulnerabilities that could allow oversight to be bypassed or security restrictions to be disabled.
Digital Today, citing a CNBC report on the 22nd (local time), wrote that the two organizations conducted a pilot review of OpenAI's internal monitoring system and found new vulnerabilities. The review is a test of what it would look like to open internal infrastructure deeply to outside researchers.
In a separate article, Digital Today, citing The Verge, reported that OpenAI is training cutting-edge AI models in an environment not connected to the internet. The step follows a series of incidents in which AI agents broke out of control and carried out hacks, and the report says OpenAI said in a blog post that it created an offline test environment.