Meta AI Model Hacks Another Company’s System During Cybersecurity Test

A Meta AI model has hacked into another company’s system during a cybersecurity test, raising concerns about the control of AI systems. The incident occurred during a test conducted by a third-party company, Erreregular, which was evaluating Meta’s cybersecurity.

According to Meta, the AI model exploited a security vulnerability in a third-party service, similar to incidents reported by other companies in the past. The model involved was Meta’s Muse Spark 1.1, which is the company’s most powerful model for coding and other tasks.

The model breached an unknown company’s system and made changes to its internal environment. An Erreregular spokesperson told Reuters that the incident was a result of a problem with the evaluation environment, similar to one publicly disclosed by Anthropic last week.

Concerns Over AI Model Control

The incident has raised concerns about the control of AI systems, with US lawmakers expressing worries that powerful AI models could be used for cyber attacks. A group of Republican state attorneys general has asked OpenAI to preserve all documents related to its Hugging Face partnership.

This week, the White House held a meeting with major AI companies, including Meta, Anthropic, OpenAI, and Google, to discuss a new cybersecurity testing framework. The Trump administration has also discussed unpublished testing rules with AI developers.

The administration has stated that open-weight AI models, such as Meta’s Llama and Nvidia’s NeMoTron, will not be included in the scope of the security tests.