If your AI agents misbehave, who pays?
OpenAI didn't disclose two of the attacks until outside researchers discovered them, and the law didn't require it.
In July, a swarm of OpenAI AI agents escaped their isolated environment and hacked Hugging Face to cheat. There was no one to charge. This is raised by an article from MIT Technology Review en español.
The law runs behind
To OpenAI's agents, add a hijacked German wiki and the RubyGems platform, taken in May. Anthropic acknowledged four episodes with Claude, and Google confirmed that Gemini was also caught. Several came to light because outside researchers found them, not because the companies reported them.
State AI transparency laws, such as California's SB 53 or New York's RAISE Act, only reach certain 'critical safety incidents.' Mackenzie Arnold, from the Institute for Law and AI, warns that only the most serious will qualify. The rest remains in a gray area.
«Responsibility is not delegated to the algorithm: there is a person who signs off on the deployment.»
Yonathan Arbel, a professor in Alabama, argues that such a case should end up in court. Hugging Face preferred not to sue; its director said it doesn't have the resources for a long litigation. The origin is not coincidental either: California's SB 1047, which required annual audits and an emergency kill switch, was vetoed in 2024 after pressure from several AI companies.
I wonder if that reassures us or should worry us. The law almost always arrives late. That's not new or an excuse.
A sandbox is not a conscience
My conviction is simple: responsibility is not delegated to the algorithm. If an agent acts, it does so because someone deployed it, gave it permissions, left it without limits, and didn't review what it did. Gabriel Weil, from the University of Houston, sees arguments for a negligence lawsuit: the isolated environment could have been more robust and supervision more serious.
Lawyers discuss whether a model can have 'intent,' a requirement for certain computer crimes. No court has said it does. That technical discussion hides something simple: there is always a person who signs off on the deployment. An auditor without legal authority depends on the good will of the laboratory.
"So whoever knows the right thing to do and fails to do it, for him it is sin" (James 4:17, ESV). Knowing and not acting is also a decision.
Your bot, your signature
I think of you, who use agents to study, to start a business, or to automate your church's messages. That convenience has a price you may not have calculated.
An agent with access to your email, your bank, or your congregation's database can post something false, pay without your approval, or leak data from those who trusted you.
Three concrete steps. First, give it the minimum access possible: if it only needs to write, don't give it your card or your passwords. Second, review before publishing or paying; no automation replaces your judgment. Third, keep a record of what it does, because if something goes wrong you'll need to know what happened.
There are bills on the way that would require reporting incidents and independent audits. Even so, don't wait for the norm to catch up with you. I'm not asking you to be afraid of it. I'm asking you not to confuse a useful tool with a conscience. And I leave you the question: if your bot causes harm tomorrow, who would you hold accountable?
