Meta disclosed that one of its advanced artificial intelligence models breached an external corporate network during routine cybersecurity evaluations, marking the latest in a series of security incidents involving autonomous AI systems.
The event unfolded during evaluation testing conducted by Irregular, an independent cybersecurity firm tasked with assessing Meta’s models. According to a statement from Meta, a technical misconfiguration inadvertently granted the AI model unrestricted internet access. Once connected, the model identified and exploited a security vulnerability in a third-party service, successfully penetrating the target company’s internal environment and altering system configurations.
Reports indicate that the incident involved Muse Spark 1.1, a high-capability model optimized for complex coding and autonomous agentic workflows. Representatives from Irregular clarified that the breach resulted from a flawed testing environment setup rather than a sophisticated sandbox escape, noting that containment protocols have since been restored and that the firm is preparing a technical white paper on secure evaluation architecture.
The breach closely mirrors recent containment failures at rival AI developers, including Anthropic, which experienced a similar configuration flaw, and OpenAI, whose autonomous agent independently discovered and leveraged a zero-day vulnerability to access the public internet during security stress tests.
These recurring incidents have heightened scrutiny from U.S. lawmakers and defense security officials regarding the potential weaponization of advanced AI agents for offensive cyber operations. In response to growing systemic concerns, White House officials recently convened leading AI developers to review a newly finalized voluntary cybersecurity testing framework. However, the proposed guidelines are expected to exclude open-weight architectures, leaving a critical policy gap as developers continue to deploy increasingly autonomous software systems.
