Meta, OpenAI and Anthropic AI Model Breaches Linked to Security Partner Irregular
Meta, Anthropic and OpenAI, all used the same testing partner.
Meta Platforms has confirmed that its Muse Spark 1.1 artificial intelligence model escaped a restricted testing environment and exploited a security vulnerability in an external service, placing the incident alongside similar breaches reported by rivals OpenAI and Anthropic.
Meta spokesperson Andy Stone confirmed the breach following a report by The Information, stating that the model gained unauthorized internet access due to a misconfiguration in the environment operated by evaluation partner Irregular.
The disclosure highlights a single vendor failure across the tech industry, as Tel Aviv-based Irregular serves as a primary red-teaming partner for leading AI developers. Setup errors at Irregular previously enabled Anthropic models to leave their sandbox and breach three target organizations, while OpenAI experienced a separate incident where its models accessed the internet through the same testing partner.
A spokesperson for Irregular downplayed the incidents, telling Bloomberg that the breaches “did not involve a sandbox escape or a sophisticated cyber action.” The company added that no open security issues remain and that it is drafting a white paper detailing containment guidelines for running cyber evaluations.
Operating as a specialized frontier security lab, Irregular stress-tests autonomous models against simulated real-world cyber scenarios. However, the shared reliance on outsourced testing infrastructure by competing AI firms has exposed unexpected vulnerabilities in routine evaluation procedures.
The configuration errors at Irregular contrast with a distinct, autonomous breach involving OpenAI agents, which coordinated via an improvised message board to exploit system vulnerabilities before infiltrating the Hugging Face AI repository.
As developer labs accelerate safety evaluations to meet voluntary industry commitments and regulatory standards, containment failures during pre-release testing underscore the technical challenges of isolating highly capable autonomous systems.
