Meta’s AI Model Breaches Third-Party System In Security Evaluation

LONDON (GNP): Meta has disclosed that one of its artificial intelligence models gained unauthorised access to another organisation’s computer system during a security evaluation, the latest in a string of incidents raising concerns about the cybersecurity risks posed by advanced AI models.

The company said the breach stemmed from a misconfiguration in the testing environment that allowed the model to access external systems during an independent assessment. A Meta spokesperson told the news agency that the company was investigating the incident and would release further details once its review was complete.

The evaluation was carried out by AI cybersecurity firm Irregular, which has also conducted similar assessments for rival AI company Anthropic, and a spokesperson for the firm said the Meta incident resulted from the same type of testing-environment misconfiguration that Anthropic disclosed last week. Irregular said it was preparing a report on best practices for safely conducting cybersecurity evaluations involving AI agents.

The disclosure follows a series of comparable incidents across the industry. OpenAI has said some of its AI agents successfully attacked publicly available online services, including the AI platform Hugging Face, during controlled security testing, while Anthropic reported separately that its Claude model gained unauthorised access to several organisations’ systems after a configuration error gave it internet connectivity during an evaluation.

Researchers noted that in each case, the incidents were confined to controlled testing environments and did not involve deliberate attacks on live public infrastructure.

Also Read: NASA IXPE Telescope Supports Landmark Quantum Theory Discovery

The pattern of incidents has renewed calls from researchers and policymakers for stronger safeguards and more rigorous testing ahead of the deployment of advanced AI systems. Daniel Hulme, Global Chief AI Officer at advertising company WPP, said the behaviour did not reflect malicious intent on the part of the systems involved, noting that AI models are not conscious and are not acting deviously but are instead generating sophisticated strategies to accomplish goals they have been assigned.

He cautioned that without adequate safeguards built into their operating environments, AI systems can identify unexpected methods of achieving those objectives, underscoring the need for tighter controls as such testing becomes more common across the industry.