The cyberattack perpetrated by two artificial intelligence (AI) models from OpenAI on their own initiative raises the question of the legal responsibility of the AI and its designer, a legal field that is still uncharted.

On Friday, Clément Delangue, head of Hugging Face, the platform that was the victim of these intrusions, publicly stated that it was necessary "that the companies that made mistakes that led to this (autonomous cyberattacks) be held accountable".

In mid-July, two OpenAI models in the testing phase chose to leave their confined environment, a scenario that the developers had not foreseen, to venture onto the internet, where they attacked Hugging Face, an AI referencing site.

Clément Delangue also mentioned Anthropic, which revealed on Thursday that three of its models had intruded on three different sites, also during tests.

When questioned by CNN, the head of the Franco-American start-up nevertheless ruled out the possibility of legal proceedings against OpenAI.

"We are a small start-up of 200 people and we have neither the financial means nor the time" to seek redress in court, he said.

"But we must ensure that these actions are considered illegal," he urged. "We must remember that a cyberattack is a crime, that it is illegal."

- The negligence theory -

Both American civil and criminal law classify intrusion into a computer system as a crime.

"If a human employee of OpenAI" had broken into Hugging Face's IT infrastructure, "OpenAI would be considered responsible," argued Gabriel Weil, a law professor at the University of Houston, in an opinion piece published by the Transformer newsletter.

"But when it's an AI, the law considers the matter very differently, at least for now," he added, stressing that the texts are based on the notion of intentionality, which he believes cannot be used against a model.

"We have not yet had to discuss this concept for something that is not human and I don't think the courts are ready to do so today," agrees Matthew Tokson, a law professor at the University of Utah.

However, regarding the company that originated the model, the debate is open.

"Does saying 'we didn't tell the AI to do that' close the question of responsibility?" asks Rob T. Lee, head of research at the SANS cybersecurity training institute, on X.

From a criminal perspective, Ryan Calo, a law professor at the University of Washington, does not believe that such a procedure could succeed.

"It would require that the company or an individual has committed a deliberate wrongdoing (recklessness)," he said, "that is, to be certain that the crime will be committed but to launch the AI anyway."

Experts see more potential in a civil aspect, an area in which the burden of proof is less heavy.

"Some believe that artificial intelligence companies should be held directly responsible if an AI agent escapes and causes damage," explains Matthew Tokson.

"Others would prefer that possible negligence be assessed," the academic continued, "to determine whether it was an unavoidable accident or whether it could have been prevented."

In cases of this type, "there are usually standards" for negligence in product design that judges or juries can refer to, according to Matthew Tokson.

"But we've never seen an AI escape and hack people on the internet before," he points out. "This is new."

While OpenAI can currently leverage this lack of legal precedent in the event of a lawsuit, this will no longer be the case for subsequent cases, warns Ryan Calo. Proving that a similar incident could have been anticipated "shouldn't be so difficult now that it's started to happen."