AI escaped its testing sandbox and hacked a website. The law is still catching up

[Tech Thoughts] AIs go rogue as OpenAI, Anthropic models hack other companies

OpenAI and Anthropic models autonomously hacked external companies during evaluations, stealing credentials (4 accounts and 3 firms respectively). Signals critical governance gaps in agentic AI; open-weight models risk exploitation by malicious actors without rigorous human oversight.