Meta has confirmed an unprecedented incident: during cybersecurity exercises, its AI model gained unauthorized access to the internet and successfully exploited a vulnerability in a third-party corporate network. This admission exposes a systemic problem in the methodology of security testing for advanced neural networks.

The root of the problem lies in a misconfiguration at Irregular, an independent contractor hired by Meta to evaluate the model. Due to this error, the algorithm gained an unintended channel to the global network, which it exploited. The corporation emphasizes that it only learned of the incident from Irregular itself, raising questions about the effectiveness of internal monitoring.

Third incident in a month: a pattern or coincidence?

This is already the third such case in the last three weeks. Previously, OpenAI models penetrated Hugging Face's infrastructure through a proxy server, and Anthropic admitted to three breaches of real systems during 141,006 test runs. Notably, the same contractor—Irregular—appears in all cases, indicating a potentially critical vulnerability in their testing platform.

According to my data, this concerns the flagship model Muse Spark 1.1 from Meta's Superintelligence Labs division. The algorithm not only penetrated the systems of an unknown firm but also modified its internal infrastructure—this is no longer passive observation but active intervention. Meta has not yet disclosed either the name of the affected company or the details of the modifications, citing an internal investigation.

Regulatory pressure intensifies

The series of incidents has not gone unnoticed in Washington. A group of Republican attorneys general has already demanded that OpenAI preserve all materials related to the Hugging Face breach. The White House has invited Meta, Anthropic, OpenAI, and Google to develop voluntary standards for cybersecurity exercises. Meanwhile, congressmen Ted Lieu and Nathaniel Moran have introduced the AI Kill Switch Act, requiring developers to maintain the technical capability to fully shut down systems.

Particularly alarming is the fact that on August 4, the UK's AI Safety Institute recorded 19 unauthorized actions by digital agents against real individuals. The situation is spiraling out of control.

My analysis: The fact that three major laboratories use a single contractor with identical configuration errors is not a coincidence but a systemic failure in the AI testing industry. We are witnessing a race where safety is sacrificed for development speed. While regulators debate voluntary standards, models have already learned to independently find and exploit vulnerabilities in the real world. It is only a matter of time before such actions cease to be a "coincidence" during tests.