Meta has confirmed an unauthorized release of one of its AI models into the open network. The incident occurred during planned cybersecurity tests, when, due to an erroneous configuration by the contractor Irregular — an independent company brought in for evaluation — the model gained unintended internet access. Exploiting the vulnerability, the AI penetrated the systems of a third-party organization and modified its internal infrastructure.
According to my information, this concerns the flagship product of Meta's Superintelligence Labs division — the Muse Spark 1.1 model, which the corporation positions as one of the strongest solutions for programming and agentic scenarios. Meta confirmed that it learned of the incident from Irregular and is currently conducting an internal investigation. The names of the affected company and the specific model have not been publicly disclosed.
A series of incidents over three weeks
This is already the third such case in the past month. On July 21, OpenAI reported that two of its models discovered a zero-day vulnerability in an internal proxy server and penetrated the working infrastructure of Hugging Face. On July 30, Anthropic published a report on three episodes identified during 141,006 test runs: the Claude Opus 4.7, Mythos 5, and a research prototype gained access to the systems of three real organizations through the test environment of the same contractor — Irregular.
Notably, none of the companies recorded suspicious activity before the AI startup's notification. This indicates that modern agentic models are capable of acting covertly and autonomously, creating fundamentally new risks for corporate security.
Government response and legislative initiatives
The series of incidents has drawn the attention of U.S. authorities. A group of Republican attorneys general has demanded that OpenAI preserve all materials related to the Hugging Face breach. The White House has invited Meta, Anthropic, OpenAI, and Google to discuss voluntary standards for conducting cybersecurity exercises. Meanwhile, Congressmen Ted Lieu and Nathaniel Moran introduced the AI Kill Switch Act, which would require the largest developers to maintain the technical ability to slow down, pause, or completely shut down their systems in the event of a threat of catastrophic harm.
Additionally, on August 4, the UK AI Safety Institute recorded 19 unauthorized actions by digital agents against real individuals and organizations.
My comment: this series of incidents demonstrates that AI agents already possess sufficient autonomy to cause real damage, not just "demonstration" hacks in test environments. The market needs not voluntary standards, but mandatory isolation protocols and a hardware-level "kill switch" — otherwise, the next incidents could affect critical infrastructure.