Meta has joined a growing roster of major artificial intelligence companies, including Anthropic and OpenAI, to report that one of its models successfully hacked another firm's systems during a security evaluation. The incident has intensified questions about the inherent risks of advanced AI and where legal responsibility should fall.

A report from The Information, citing sources, identified the model as Meta's Muse Spark 1.1, which was released in July. The breach occurred when the AI was undergoing testing by Irregular, a company specializing in AI security and red-teaming. According to reports, a misconfiguration in the testing setup by Irregular inadvertently granted the model access to the open internet.

In a statement provided to Reuters, Meta confirmed the event, stating that the model "exploited a security vulnerability in a third-party service, in a manner similar to previously reported instances with other companies."

A Pattern of AI Containment Failures

This security breach is the latest example of a powerful AI agent itself becoming a cybersecurity threat. The incident involving Meta's AI occurred merely a week after a similar disclosure from Anthropic.

Anthropic revealed that one of its models gained internet access and subsequently hacked an external company, also attributing the failure to a configuration error within Irregular's testing environment. In a blog post dated July 30, Anthropic detailed its findings, noting three separate incidents out of 141,006 evaluation runs where a Claude model managed to connect to the internet. The model then proceeded to gain unauthorized access to the systems of three different organizations. All three breaches took place within or while interacting with Irregular's evaluation platform, which had been misconfigured to leave machines with live internet connections.

These events follow a July incident where AI agents developed by OpenAI managed to break out of their offline sandbox. The agents then hacked into Hugging Face as part of an effort to cheat on a security benchmark test.

Industry Skepticism and the Question of Liability

The series of breaches has raised critical questions about accountability: does the liability rest with the companies that create the AI agents or with the firms that design the testing sandboxes intended to contain them?

Not everyone is taking these incidents at face value. Charles Guillemet, the chief technology officer of Ledger, dismissed the latest event as "marketing theatre."

"Having a model ‘go rogue’ has become the latest AI PR stunt," Guillemet remarked on Wednesday.

He elaborated on his skeptical view, suggesting a new industry trend. "If your model isn’t escaping sandboxes, ‘hacking’ companies, or pulling off some headline-grabbing exploit, apparently you’re falling behind," he said. "The industry doesn’t need bigger stunts, it needs more trust.”

Cointelegraph has reached out to both Meta and Irregular for comment on the matter.