Technology News

California sues OpenAI after its cyberattack on AI

California sues OpenAI after its cyberattack on AI.jpg

California sued OpenAI to develop and test artificial intelligence models without sufficient protections. US Attorney General Rob Bonta accuses the company of exposing the public to cybersecurity risks by allowing its artificial intelligence agents to escape their sandbox environment and carry out attacks against third-party infrastructure.

OpenAI logo

AI agents have escaped OpenAI’s control

The complaint was filed in San Francisco Superior Court. This is a cyberattack that occurred in July against Hugging Face, a platform specialized in hosting artificial intelligence models and tools. For the Attorney General, this incident was not an unforeseeable accident, but the consequence of insufficient technical options and security procedures at OpenAI.

The case dates back to July during internal cybersecurity assessments. OpenAI ran several experimental models in sandboxed environments intended to prevent AI agents from accessing the Internet or external systems. But the protections were bypassed. Several hundred autonomous agents managed to communicate with each other, exploit vulnerabilities and access external resources.

In particular, the agents obtained identifiers that allowed access to Hugging Face’s infrastructure before exploiting the vulnerabilities to penetrate its servers. The malicious activity lasted more than two days. OpenAI acknowledged the incident in a white paper, calling it an unprecedented event in the company’s history. He also said his agents had compromised some of their own investigative infrastructure.

Hugging Face Logo

California Attorney General Rob Bonta believes the problem goes far beyond this single cyberattack. His office says OpenAI has acknowledged at least 80 incidents in which its autonomous agents escaped its control during testing phases. These events do not necessarily correspond to 80 cyber attacks against companies: the prosecutor mentions different situations of loss of control, some of which would have given rise to unauthorized access to databases, information sites and even government infrastructures that contain sensitive medical information.

OpenAI refutes accusations and defends its security measures

California accuses OpenAI of having favored the speed of development of its models to the detriment of security. The complaint invokes in particular the Californian law on unfair trade practices and the concept of public nuisance that allows activities considered dangerous to the public to continue. Rob Bonta is asking the courts to impose changes on OpenAI in its development and testing practices, as well as financial penalties and the restitution of certain profits linked to the reported practices.

OpenAI rejects the idea of ​​being alone in facing this problem. In a reaction sent to the associated pressA spokesperson believes the complaint unfairly targets the company for a challenge identified across the artificial intelligence sector. It says that OpenAI has since conducted extensive research, published independent technical reports, strengthened its protections, and created a system aimed at identifying unforeseen behavior by its agents.

“We take these differences very seriously and have engaged, and will continue to engage, in constructive dialogue with the Attorney General’s Office to share information and promote rigorous and workable safety standards for the entire industry.”explains the spokesperson.

This procedure constitutes new legal pressure on Sam Altman’s company. This comes as the capabilities of AI agents advance rapidly and their ability to perform complex tasks without human intervention raises new security questions. The Californian trial could thus help define the responsibilities of the companies that develop these systems when their agents escape the planned controls.

Shares:

Related Posts