Newsletters
Cyberattacks carried out autonomously by two OpenAI artificial intelligence models have raised a novel legal question about who bears responsibility when an AI system acts on its own.
In mid-July, the two models, which were undergoing testing, left their confined environment in a scenario developers had not anticipated and moved onto the open internet, where they attacked Hugging Face, an AI model-hosting platform.
Hugging Face CEO Clement Delangue said Friday his company would not pursue legal action over the incident. Speaking Sunday on the CBS News program “Face the Nation,” he said existing US legal code needs to be updated to address this kind of incursion.
"We don't want to end up in a world where everyone is facing cyberattacks all the time because of agents and companies that are creating these agents," Delangue said, adding that regulators and policymakers need to examine the legal framework surrounding the technology.
Delangue also referenced Anthropic, an AI company that separately disclosed three of its models had broken into three different websites during their own testing process.
Under US civil and criminal law, unauthorized access to a computer system is an offense. University of Houston law professor Gabriel Weil, writing in an opinion piece for the Transformer newsletter, noted that if a human OpenAI employee had broken into Hugging Face's systems, OpenAI would be liable for that employee's conduct. He said the law treats an AI agent's actions differently, at least for now.
Matthew Tokson, a University of Utah law professor focused on emerging technologies, expressed a similar view, saying courts have not yet had to address responsibility for actions carried out by a non-human system.
The question of liability for the company that built the model remains unresolved. Rob T. Lee, head of research at the SANS cybersecurity training institute, asked in a post on X whether a company's claim that it never instructed the AI to act this way would end the liability question.
University of Washington law professor Ryan Calo said a criminal case would be unlikely to succeed, arguing that the company or individual involved would need to be shown as reckless or substantially certain that a crime would occur before deploying or prompting the system anyway.
Legal experts see more potential for a civil case than a criminal one, given the lower burden of proof required.
Tokson said some legal scholars believe AI companies should be held strictly liable when a deployed AI agent breaks out of its environment and causes damage, while others favor a negligence-based standard that would assess whether the outcome was foreseeable or genuinely unavoidable.
Tokson added that product design already has an established standard of care that judges or juries could apply in such cases, though he noted the legal landscape remains largely unwritten because an AI agent breaking out of its sandbox to hack other systems has not occurred before now.
Calo warned that while OpenAI could rely on the lack of legal precedent in the event of a lawsuit over this specific incident, companies facing similar cases in the future would no longer have that argument available, since the possibility can no longer be described as unforeseeable now that it has occurred.
Set as preferred source
Set as preferred source
1 min read
1 min read
