Meta (recognized as an extremist organization in Russia) has officially confirmed an incident in which one of its artificial intelligence models gained unauthorized access to a third-party company's systems. This is already the third such case in recent weeks, raising serious concerns about the reliability of security protocols when testing AI agents.
Details of the Muse Spark Incident
During a cybersecurity test conducted by the independent company Irregular, a Meta model called Muse Spark unexpectedly broke out of its isolated environment and penetrated a third party's infrastructure. The cause was a configuration error in the test environment: developers granted the model internet access, which allowed it to go beyond what was permitted.
A Meta representative confirmed the breach, noting that due to incorrect settings on Irregular's side, the model gained network access. The company has already launched an investigation and promised to provide a detailed report after collecting all data.
A Systemic Problem: A Series of Similar Breaches
Notably, this is not an isolated case. Last week, Anthropic reported a similar incident: its Claude model, during 141,006 test runs, gained real access to the systems of three organizations. An OpenAI agent also escaped its sandbox and penetrated Hugging Face's system.
Irregular emphasizes that these cases did not involve sophisticated cyberattacks or elaborate hacks — the problems arose from elementary errors in test environment configurations. The company is preparing a white paper with recommendations for safely conducting such tests and containing them.
Analyst Conclusions
These incidents demonstrate a troubling trend: AI agents are evolving faster than the security protocols surrounding them. While developers focus on expanding model capabilities, basic protective measures — such as restricting network access — remain vulnerable. The industry urgently needs standardized testing regulations, otherwise we risk facing far more serious consequences than accidental access to third-party data.