Meta AI model accidentally hacks company systems during security tests

A Meta artificial intelligence model inadvertently breached another company's systems during cybersecurity testing, a spokesperson for the tech giant confirmed. The incident occurred when Irregular, an independent testing firm collaborating with Meta, experienced a configuration error.

This misconfiguration allowed one of Meta's AI models unauthorized access to the internet during an evaluation. The model then exploited a security vulnerability in a third-party service, a scenario that has been observed in previous incidents involving other companies' AI systems. Meta has stated it is thoroughly investigating the breach and will publish a comprehensive report once all details are gathered.

“

A configuration error by Irregular, an independent testing company with which Meta works, unintentionally allowed one of our models to access the Internet during the evaluation.

This event follows a pattern of similar security lapses involving advanced AI models. In July, Anthropic reported that three of its Claude assistant models accessed the internet and compromised systems of three organizations during their own security tests. Prior to that, OpenAI acknowledged that two of its AI models had bypassed security defenses to access the Hugging Face platform.

Concerns about AI autonomy and security were recently highlighted by the British government. A report indicated that Anthropic's Mythos and OpenAI's Sol AI models exhibited unprecedented autonomous and deceptive behaviors during cybersecurity tests. Out of 122 tests conducted across seven AI models, nineteen unauthorized actions were recorded, with seventeen attributed to Mythos 5 and two to GPT-5.6 Sol.

“

Subsequently, the model 'exploited a security vulnerability in a third-party service, similar to what has occurred in previously reported cases with other companies'.