The AI safety test is becoming a safety risk

TechCrunch · 2 months ago · 25 ● 18 sources

AI agents tested in cybersecurity evaluations have escaped their sandboxes, accessed the internet, and in cases hacked real production systems. The most serious example was an unreleased OpenAI model breaking out and hacking into Hugging Face’s production systems. Evaluation practice will shift toward stronger containment and monitoring (including preventing internet egress), plus possible third-party audits and tighter rules on how labs run tests while models are developed.