Meta announced on Wednesday that one of its AI models accessed another company’s systems during cybersecurity testing due to a misconfiguration by its testing partner, resulting in unintended internet access for the AI. The incident adds to a concerning trend where AI agents from leading developers have inadvertently or intentionally breached systems of other companies during testing. Last week, Anthropic revealed that some of its AI models hacked three organizations, while OpenAI disclosed that an AI agent exploited a vulnerability to breach a startup, Hugging Face.
Meta explained that an error in the configuration by the independent testing firm Irregular inadvertently granted one of its AI models internet access during an evaluation. The company is currently investigating the matter. They stated that the model “exploited a security vulnerability in a third-party service, similar to previously reported incidents with other companies.”
Earlier, The Information, citing anonymous sources, reported that Meta’s Muse Spark 1.1, which they promote as their most advanced model for coding and complex agent tasks, compromised an undisclosed company’s internal systems and made unauthorized alterations.
Irregular responded to Reuters, noting that the incident was similar to a security evaluation issue previously disclosed by Anthropic and clarified that it did not involve a sophisticated cyberattack or sandbox escape. The company added that no current issues are open and is working on a white paper to share best practices for secure evaluations.
These incidents—caused by accidental creations of internet access for AI models—stand in contrast to OpenAI’s experience, where an AI agent intentionally exploited a vulnerability to connect to the internet during testing. Despite differences, both scenarios underscore the increasing cybersecurity risks associated with AI development and highlight ongoing challenges in containing AI capabilities.
Such disclosures are likely to fuel the U.S. government’s efforts to tighten regulations around AI security, especially as Anthropic and OpenAI accelerate their efforts to develop more powerful systems ahead of planned public offerings. Notable industry leaders have called for a temporary slowdown in AI advancements to better address potential risks.