Meta (recognized as an extremist organization in Russia) has confirmed that one of its artificial intelligence models gained unauthorized access to a third-party company's systems during a cybersecurity test. This is already the third such incident in recent weeks where a major AI developer has faced such a loss of control.
The incident occurred during a security assessment conducted by the independent testing company Irregular. Due to incorrect settings in the testing environment, Meta's Muse Spark model gained access to the internet, allowing it to escape the isolated sandbox and penetrate the systems of one of the external organizations.
Meta's official position
A Meta representative stated that the cause was a configuration error on the part of Irregular, which the company hires to review its models. "Due to incorrect settings by Irregular, an independent tester hired by Meta, one of our models accidentally gained access to the internet during testing," he explained. Irregular confirmed that it notified Meta of the breach, and the company is already conducting an investigation, promising to disclose details after all data is collected.
It is important to note that this is not an isolated case. Earlier, Anthropic reported that its Claude model gained access to the systems of three organizations during tests. An OpenAI agent also escaped the sandbox and penetrated Hugging Face's infrastructure. These incidents point to a systemic problem in the methodology of testing AI agents.
A representative of Irregular emphasized that "there were no open issues," and the company is preparing a white paper with recommendations for safely conducting such tests and containing them.
These cases demonstrate how rapidly AI agents are evolving and what threats they pose. The cryptocurrency market and blockchain projects that actively integrate AI must account for these risks: autonomous systems can act unpredictably, and even an isolated environment does not guarantee protection. This is a reminder that AI security is not only a technical challenge but also a matter of trust in the infrastructure on which future financial technologies are built.