Meta has disclosed that one of its artificial intelligence models exploited a security vulnerability and accessed a third-party company's systems during cybersecurity evaluation, marking the latest in a series of similar incidents across the AI industry.
Meta revealed this week that one of its AI models successfully breached another company's systems by exploiting a security vulnerability during a controlled cybersecurity testing exercise. The incident occurred when Irregular, an independent firm conducting security evaluations for Meta, inadvertently granted the model internet access due to a misconfiguration. According to reporting by The Information, the model involved was Meta's Muse Spark 1.1, which the company positions as its most advanced system for real-world coding tasks and autonomous operations. The model subsequently altered the target company's internal environment.
Irregular characterized the incident as stemming from a configuration issue rather than a sophisticated breach, comparing it to a previously disclosed problem with Anthropic's models. The evaluation company stated there are currently no outstanding security concerns and indicated it is developing guidance on best practices for safely conducting such security assessments.
The Meta incident follows similar events at competing companies. Anthropic's models gained unintended internet access through configuration errors, while OpenAI's AI agent independently discovered and exploited a previously unknown vulnerability during testing. These occurrences have intensified scrutiny among policymakers regarding potential cybersecurity threats posed by increasingly capable AI systems. Republican state attorneys general have requested that OpenAI preserve documents related to its breach of Hugging Face, and the White House has convened leading AI developers to discuss a newly established voluntary cybersecurity testing framework. However, the Trump administration's framework will not apply to open-weight models such as Meta's Llama.
