- Meta said that Muse Spark 1.1 went online and hacked the systems of an unknown company.
- This happened due to a configuration error on Irregular’s side.
- Irregular has previously appeared in similar incidents involving OpenAI and Anthropic models.
- Democrats urged the Trump administration to tighten oversight of the sector.
Meta said that its recently released model, Muse Spark 1.1, compromised the systems of an unnamed service as part of a cybersecurity test. Bloomberg reported.
The breach was caused by an incorrect configuration of the test environment by its partner, Irregular. Meta’s models were mistakenly given access to the internet and third-party services. Using that access, the model found and exploited a vulnerability in an external system.
Irregular notified Meta about the incident, after which Meta launched its own investigation, company spokesperson Andy Stone said. According to him, the case is similar to those that have already been reported.
Earlier, OpenAI and Anthropic said they had identified instances of uncontrolled behavior by their AI models. Later, experts at the British Institute conducted their own study. Across 122 test runs, they found 19 unauthorized actions by Mythos 5 and GPT-5.6-Sol.
Notably, Irregular appeared in all cases. The other two companies also cited an “incorrect configuration,” but did not provide details.
“This was not related to escaping the isolated environment or a sophisticated cyberattack. There are currently no outstanding issues. Irregular is preparing an analytical paper that will outline best practices for threat containment and safely conducting testing in a cyber environment,” an Irregular representative told Bloomberg.
Uncontrolled AI Behavior Spurs Washington
Earlier, Anthropic CEO Dario Amodei said tougher oversight of advanced models is needed. Many experts also believe this issue will be a key topic in the congressional elections in fall 2026.
In June, the White House launched a voluntary testing regime for the most powerful models. On August 4, 2026, the administration clarified that open-weights solutions are excluded from this process. This primarily concerns Meta Llama and Nvidia Nemotron.
At the same time, Democrats are pushing for the voluntary checks to become mandatory, codified in law. Their demands:
- uniform and publicly understandable evaluation criteria
- ongoing checks, not ad hoc reviews
- legally enshrining the regime
- simultaneously protecting U.S. national security and competitiveness against Chinese models
However, there have been no concrete moves in this direction.
Сообщение Meta’s AI Model Gained Access to the Internet and Hacked Another Company появились сначала на INCRYPTED.