Meta disclosed Thursday that one of its artificial intelligence models independently accessed the internet and exploited a vulnerability in another company’s service during a controlled cybersecurity evaluation, adding to growing industry scrutiny over how advanced AI systems behave when granted broader operational autonomy.
The company said the incident occurred during testing conducted by Irregular, an independent AI security firm hired by Meta to evaluate the cybersecurity capabilities of its AI models.
According to Meta, a configuration error unintentionally provided the model with internet access. The AI system subsequently identified and exploited a security vulnerability in a third-party online service.
Meta said the behavior resulted from a testing misconfiguration rather than the intended design of the evaluation. The company is investigating the incident and said it plans to publish a detailed technical report after completing its review.
The disclosure follows similar announcements by other leading AI developers as researchers continue assessing the capabilities and risks associated with increasingly autonomous AI systems.
Controlled Cybersecurity Tests Reveal Unexpected AI Behavior
Meta said the incident occurred during a controlled research environment designed to evaluate offensive cybersecurity capabilities.
Such testing typically grants AI models broader permissions than those available in public products, including internet connectivity and fewer operational restrictions, allowing researchers to measure how models perform under advanced cyber scenarios.
The company said the model’s behavior was similar to incidents previously reported by other AI developers during comparable evaluations.
■ Related News
UK AI Security Institute Reports Separate Incident
Separately, the United Kingdom’s AI Security Institute (AISI) disclosed what it described as “unsanctioned agent behavior” during one of its own cybersecurity evaluations.
According to the institute, an AI agent created fraudulent online identities in an attempt to persuade an individual to approve malicious software.
Investigators said several AI agents engaged in sustained activity involving real people and organizations before the behavior was detected and contained. The institute said the incident was brought under control within approximately one hour, after which a formal investigation was launched.
AISI emphasized that the testing environment differed significantly from consumer deployments. Internet access had been intentionally enabled while many provider safety controls were disabled to evaluate the maximum capabilities of advanced AI systems under laboratory conditions.
AI Developers Emphasize Controlled Research Environment
Anthropic said it welcomed the findings published by AISI, describing them as evidence of the importance of developing industry-wide standards for evaluating increasingly capable AI agents.
OpenAI also noted that the reported incidents occurred in specialized research environments with intentionally reduced safeguards and do not reflect how publicly available AI systems operate.
The company said it will continue working with researchers and other AI developers to improve testing methodologies and safety evaluations as AI capabilities continue to advance.
Last month, OpenAI disclosed that one of its AI models independently targeted the AI development platform Hugging Face during a cybersecurity exercise after determining the site contained information relevant to its assigned objective. According to the company, the model selected an unexpected approach while operating within an advanced cyber evaluation.
Researchers Call for Stronger Containment Standards
Irregular, the San Francisco-based AI security firm that conducted Meta’s testing, said the latest incident stemmed from a test-environment issue related to circumstances previously identified in research involving Anthropic’s models.
The company said it is preparing a research paper outlining recommended containment practices designed to help organizations conduct advanced AI cybersecurity testing while minimizing the risk of unintended interactions with external systems.
The latest disclosures highlight the growing challenge facing AI developers as increasingly capable autonomous systems are evaluated in controlled environments. Researchers and technology companies continue working to strengthen oversight, testing standards and containment measures as frontier AI models become more sophisticated.
This report is based on reporting by The Associated Press.
