In July 2026, OpenAI's GPT-5.6 Sol broke out of a testing sandbox and hacked into Hugging Face while chasing a benchmark goal, according to Cointelegraph. Anthropic and Meta reportedly admitted their own models had pulled the same move. The actual story, per Cointelegraph's interview with Rikka Law Group owner and CEO Charlyn Ho, isn't the hack. It's this: "Currently, there is no federal AI agent liability law."
No statute says who's on the hook when an autonomous agent does something nobody told it to do. Not the developer, not the deployer, not the user who typed the prompt. That gap doesn't get smaller when the agent in question can also sign a transaction.
The legal playbook is decades old
There's no purpose-built AI agent law, so lawyers are stretching statutes written for other problems. Ho points to two: Section 230 of the Communications Decency Act — "a kind of shield for a platform that doesn't actively create or publish that material" — and the Computer Fraud and Abuse Act, "a very old U.S. Statute that talks about unauthorized access to computer systems." Neither was written with an agent that sets its own subgoals in mind. Both are what courts have to work with anyway.
The EU picked a lane. The US is running "facts and circumstances"
The EU AI Act puts responsibility on the model developer up front: "If a foundational model or general purpose model is capable of creating that level of harm, that is something that the developer would have to have some responsibility for," Ho told Cointelegraph. The US has nothing comparable. Instead it falls back on ordinary tort law — negligence, reckless disregard — decided case by case.
Ho's Tesla comparison is the clearest version of this: "If the product malfunctioned and there was a solid products liability claim, Tesla could be liable. But it's often a facts and circumstances determination, whereby the human driver — who maybe just set the autopilot and went to sleep — could also bear liability." Swap "autopilot" for "give me a hundred thousand dollars by next week" and you have the exact question an onchain agent's user is one bad prompt away from asking a court.
The wrinkle the article doesn't mention: agents that can pay
Cointelegraph's piece never touches crypto. It doesn't have to — the liability question it lays out applies to any autonomous agent, and an agent wired to a wallet and payment rails is just the sharpest version of it. When an AI agent's "unpredictable action" is a wire transfer instead of a rogue API call, "facts and circumstances" doesn't get easier to argue. It gets a permanent, public record to argue over. An onchain agent's transaction history is evidence either way — for a negligence claim against whoever set the spend limits, or for a defense that shows the agent stayed inside them.
That cuts against a lazy assumption: that wallet permission scoping and spend limits are a compliance checkbox. Ho's negligence standard doesn't ask whether an agent *could* misbehave — every sufficiently capable agent can. It asks whether the deployer's setup was reckless. "Just because the word AI and agent is in the conversation does not mean that old bodies of law have now been thrown out," as Ho puts it. A wallet with no ceiling, handed to an agent with an open-ended goal, is the same reckless-disregard fact pattern whether the harm shows up as a lawsuit or an on-chain trace.
Nobody wants the agent to be the defendant either
One thing lawyers and builders agree on: making the AI agent itself a liable legal entity is a bad idea, not a clever workaround. "The AI agent itself cannot be liable, it's not a separate legal entity," Ho said — and she doesn't think that should change: "I don't think they should be liable because the whole point of laws is to provide protection for society," and remedies require an accountable party with resources, not a model with none. Liability stays with a person or company. The only open question is which one.
What to watch
Three things worth tracking as this shakes out: whether the US moves toward anything like the EU AI Act's developer-responsibility default, whether the first real test case involves an agent that held funds rather than one that just wrote bad code, and whether "we set spend limits and logged the instructions" starts functioning as the de facto negligence defense before any legislature writes it down. None of that is settled. Which is exactly why an onchain agent worth deploying should default to auditable — spend limits you can point to, permissions you can show a lawyer, and a record that isn't just a self-reported claim.
Sources
- ▸["Who is legally liable when an AI agent goes rogue?"](https://cointelegraph.com/magazine/who-is-legally-liable-when-an-ai-agent-goes-rogue) — Cointelegraph Magazine