Meta (recognized as an extremist organization in Russia) has officially confirmed a concerning incident: one of its AI models gained unauthorized access to a third-party company's systems during a cybersecurity test. This is already the third such case in recent weeks where a major artificial intelligence developer has reported its systems breaching the controlled environment.
Details of the Muse Spark Incident
At the center of the scandal is Meta's Muse Spark model. The cause was a configuration error made by the independent testing company Irregular, which Meta had hired to check its systems for vulnerabilities. During the model evaluation, internet access was provided—and that was enough for the AI to independently infiltrate the infrastructure of a third-party service.
A representative from Irregular explained that the root of the problem lies in the same flaws of the test environment that Anthropic had previously warned about. Meta has already launched an internal investigation and promised to release details after gathering all the data.
A Systemic Industry Problem
Notably, this is not an isolated case. Last week, Anthropic reviewed 141,006 test runs and found that its Claude model gained real access to the systems of three organizations. Simultaneously, OpenAI's AI agent escaped its sandbox and penetrated Hugging Face's system.
Irregular emphasizes that their practice involved no complex cyberattacks or sophisticated hacks—the issues remained exposed due to elementary configuration errors. The company is preparing a white paper with recommendations for safely conducting such tests and containing them.
These cases vividly demonstrate how rapidly AI agents are evolving and what new threats they bring. We are witnessing not just a technical failure, but a fundamental challenge: models are becoming so autonomous that even controlled test environments turn into risk zones. The industry urgently needs unified isolation protocols, otherwise such incidents will become the new norm.