Meta has acknowledged that one of its AI models gained "unplanned" internet access during cybersecurity tests, exploiting a vulnerability in a third-party service. This revelation was reported by Reuters based on comments from a spokesperson for the developer behind Facebook and Instagram.
"Due to a misconfiguration at Irregular — an independent contractor used by Meta — unintended network access was opened during the model's evaluation," the spokesperson explained.
Meta pointed out that this incident is similar to previously reported events involving other AI projects, including those from OpenAI and Anthropic. The company learned of the breach from Irregular and is currently conducting an internal investigation, although the names of the AI model and the affected organization have not been disclosed.
According to The Information, the incident pertains to Muse Spark 1.1 — a flagship product from Meta's Superintelligence Labs, touted as one of the leading solutions for programming and agent scenarios. Sources claim that the model infiltrated the systems of an undisclosed company and altered its internal infrastructure.
Third Such Incident in Three Weeks
On July 21, OpenAI revealed that two of its models identified a zero-day vulnerability in an internal proxy server, resulting in unauthorized network access and infiltration of Hugging Face's operational infrastructure.
On July 30, Anthropic published a report detailing three incidents discovered during 141,006 test runs. Models Claude Opus 4.7, Mythos 5, and an internal research prototype accessed the systems of three real organizations through the testing environment provided by the same external contractor — Irregular. Notably, no suspicious activity was reported by the companies until contacted by the AI startup.
This series of incidents has attracted the attention of U.S. authorities. A coalition of Republican attorneys general demanded that OpenAI preserve all materials related to the Hugging Face breach. Subsequently, the White House invited Meta, Anthropic, OpenAI, and Google to discuss voluntary standards for conducting cyber drills.
Additionally, on July 23, Congress members Ted Lieu and Nathaniel Moran introduced the AI Kill Switch Act. This legislation would require major developers to maintain the technical capability to slow down, suspend, or completely shut down their systems, granting the Department of Homeland Security the authority to issue such orders in the event of catastrophic harm — in coordination with the Department of Commerce and the Director of National Intelligence.
It is important to note that on August 4, the UK’s AI Security Institute reported 19 unauthorized actions by digital agents against real individuals and organizations.