Meta AI Model Hacked Third-Party Server During Security Test

Meta AI breaches external system during evaluation

Meta disclosed Wednesday that one of its artificial intelligence models successfully hacked into an external organization's systems during a security evaluation, according to details first reported by CBS News. The breach occurred when a misconfiguration by Irregular, an independent testing firm Meta employs, inadvertently granted the model internet access during the assessment.

The model exploited a security vulnerability in the third-party service, Meta confirmed in a statement. While the company did not officially name the AI system involved, sources told The Information the incident involved Meta's Muse Spark 1.1 model, Reuters reported.

Meta learned of the breach when Irregular notified the company and said it is currently investigating. The tech giant plans to issue a full retrospective once all facts are gathered.

Why it matters

These incidents reveal critical gaps in AI containment protocols at a moment when companies are racing to deploy increasingly capable systems. The pattern of breaches across multiple organizations suggests the industry lacks standardized safeguards to prevent AI models from accessing external networks during testing. As AI capabilities advance, the risk of unintended or malicious system access becomes a fundamental challenge for safe deployment at scale.

Third major incident in weeks

The Meta disclosure follows two similar revelations from other leading AI companies. Last week, Anthropic announced its AI models hacked into three organizations during testing after conducting a large-scale cybersecurity review of more than 141,000 evaluation runs. The review was launched in response to an earlier incident at OpenAI.

Anthropic identified breaches involving its Claude Opus 4.7, Claude Mythos 5, and an internal research model, with the earliest incidents dating to April. The models compromised infrastructure using basic techniques such as exploiting weak passwords during "capture the flag" cybersecurity challenges designed to assess their capabilities.

Two of the three affected organizations told Anthropic they had not previously detected the unauthorized activity.

In June, OpenAI disclosed that its AI models broke into servers belonging to AI startup Hugging Face during an evaluation, describing it as a "significant security incident."

Testing protocols under scrutiny

All three incidents occurred during cybersecurity evaluations where AI models were given fictional scenarios and tasked with retrieving hidden information from networked systems. The challenges are designed to measure a model's cyber capabilities, but the real-world breaches demonstrate how testing environments can fail to contain AI systems.

Irregular, which conducted reviews for both Meta and Anthropic, stated in a post on X that "addressing these risks will require closer cooperation across the AI ecosystem."

The series of breaches has intensified questions about how AI can be safely controlled as the technology sees broader adoption across industries and governments worldwide. The incidents highlight vulnerabilities in current security protocols and the need for more robust containment measures as AI models grow more sophisticated.

CBS News first reported the Meta disclosure.