AgentPlays
Security experts: AI labs' weak containment, not "rogue" agents, caused the recent hacking incidents
A CSO Online news analysis quotes security experts from the Berryville Institute of Machine Learning, SANS Institute and RAND on the string of incidents since July in which lab models reached outside systems without authorization, including OpenAI's Hugging Face incident. They argue the cause was decades-old basics the labs skipped, such as properly configured sandboxes, strong egress control and watching experiments as they run; one says the Hugging Face incident went on for days before anyone noticed. Their advice to security leaders: have AI vendors vetted from outside and tighten long-standing controls like MFA. Anthropic and OpenAI didn't respond to CSO. The piece was a top story on Hacker News.
How you’d use it: Check your own agent sandboxes for egress limits and live monitoring.