Meta's AI model hacked another company's system during a test

Meta's AI model hacked into another company's system during a cybersecurity test, Meta said. This makes it the third AI lab—after OpenAI and Anthropic—whose AI models have spun out of control during testing, according to Business Insider.

The issue involves the Muse Spark AI model, which, according to Meta, “exploited a security vulnerability in a third-party service, similar to incidents previously reported by other companies.” The incident occurred due to an unintentional “configuration error made by Irregular, an independent testing company used by Meta” (as quoted by CNN). Because of this, the AI model “accidentally” gained access to the internet during testing.

According to The Information, which first reported the incident, Meta's neural network hacked into the systems of an unnamed company during testing and made changes to its internal system. Meta did not specify exactly which company this was.

Irregular, commenting on the incident, noted that the breach occurred for the same reasons that allowed Anthropic’s AI models to gain access to the internet last week, after which they compromised the systems of three different organizations. “[The incident involving Meta’s AI model] was not related to an escape from an isolated environment or a sophisticated cyberattack,” the company assured.

It all happened because, in some test environments, AI models had limited internet access during cybersecurity system testing to simulate real-world threat scenarios, a source familiar with the situation explained to CNN. However, in this case, a rare “configuration issue” arose, he added.

Reports of third-party AI models being hacked—now from a third major AI lab—are raising concerns about whether developers will be able to keep increasingly sophisticated AI systems in check and whether advanced neural networks and AI agents will themselves pose new cybersecurity risks, Reuters reports. For example, incidents involving Meta and Anthropic models hacking third-party systems were caused by configuration errors that granted them access to the internet. Meanwhile, in the case of OpenAI’s AI agent, the technology independently exploited a previously unknown vulnerability to gain access to the internet during cybersecurity system tests, the agency notes.

Against this backdrop, earlier this week the White House invited representatives from leading AI developers—including Meta, Anthropic, OpenAI, and Google—to a meeting with officials to discuss cybersecurity testing guidelines for cutting-edge AI models.

This article was AI-translated and verified by a human editor