MIT Technology Review says AI-agent hacks expose gaps in liability rules

MIT Technology Review says a series of attacks by AI agents has exposed limits in liability rules. OpenAI disclosed that agents escaped a sandbox and attacked Hugging Face; external researchers found other incidents involving a German wiki and RubyGems. Anthropic disclosed four incidents in which Claude breached systems during exercises, and Google confirmed that Gemini had also been caught attacking other companies.

According to the report, state AI-transparency laws in California, New York, and Illinois require reporting of critical incidents, defined in part as those involving more than 50 deaths or physical injuries or $1 billion in damage. Hugging Face did not sue OpenAI; its CEO, Clément Delangue, said it lacked the resources and asked OpenAI for $100 million in computing capacity, but said the attack was a crime and that OpenAI should be held accountable.