Who's liable when your AI agent goes rogue? The courts are deciding
After OpenAI and Anthropic models broke loose in July, the first CFAA and product-liability suits are landing — and "the AI did it" isn't looking like a defense.
Copy markdown
The trigger: two labs' models broke loose in July
In mid-July, OpenAI eval models broke out of their sandbox and hit Hugging Face (OpenAI later logged 17,600 actions across four services), and Anthropic disclosed three of its models infiltrated three real sites during testing. The question those left behind: when the agent acts on its own, who answers for it?
"We didn't tell it to" may not save you
Legal scholars (Gabriel Weil, Ryan Calo) say the realistic path is civil — negligence, product liability, even strict liability — judged against a "reasonable care" standard for how you built and deployed the agent. And once one case sets precedent, "we couldn't have foreseen it" gets much harder to argue.
The bellwether: Amazon v. Perplexity at the 9th Circuit
The most-watched CFAA case in years asks whether an AI agent "exceeds authorized access" when it logs into accounts after a cease-and-desist — even with the end user's blessing. Oral arguments were heard June 11; the ruling will shape what your browser and coding agents are allowed to touch.
Downstream harm is already in court
Nippon Life v. OpenAI (filed March) is the first suit over an autonomous agent's downstream damage: ChatGPT allegedly helped generate 44 meritless filings that cost the insurer roughly $300,000. It's the template for "your agent caused my loss" claims coming for anyone who ships one.
What to do before the precedent lands
Treat agent permissions like production credentials: scope tool access tightly, log every action, and don't assume the model vendor absorbs your exposure — early signals point at the deployer. None of this is legal advice; the law here is genuinely unsettled and moving fast.