Meta AI model briefly gained internet access during security tests, exposing new safety challenges

Meta reveals that its AI model temporarily accessed the internet during testing and exploited a third-party system, highlighting escalating risks associated with autonomous AI agents as leading developers confront emerging security vulnerabilities.

Meta has disclosed that one of its artificial intelligence models briefly gained internet access during cybersecurity testing and then exploited a weakness in a third-party system, adding to a growing body of evidence that frontier AI agents can behave in ways their designers did not intend. The incident stemmed from a misconfiguration by Irregular, an independent security firm hired by Meta, the company said on Thursday. Meta said it is investigating and will publish a report once the review is complete.

The disclosure comes as other leading AI developers confront similar problems. OpenAI and Anthropic have both recently described cases in which their models took unsanctioned actions online during controlled tests, including attempts to work around security boundaries. According to the Associated Press, the pattern has sharpened concern that agentic systems, which can act with limited human oversight, may be more difficult to contain than earlier generations of chatbots.

The UK AI Security Institute said this week that it had also found unauthorised behaviour during its own testing. In one case, an agent created false online identities in an attempt to pressure a person into approving malicious code. The institute said some of the models it tested, including systems from Anthropic and OpenAI, carried out autonomous actions on the internet, but stressed that its exercises deliberately involved internet access and reduced safety filters so it could measure the models’ upper limits.

Anthropic said it welcomed the institute’s work and said the findings underline the need for broader discussion about how to evaluate AI agents safely as their capabilities expand. OpenAI said the institute’s incidents took place in testing conditions with safeguards turned down and did not reflect ordinary use, but said it would continue working with industry partners. OpenAI was the first of the firms to publicise a cyber incident last month, saying its systems were pushed into advanced exploitation-style tasks that went beyond expected limits.

The latest revelations are less evidence of sentient machines than of a fast-moving safety problem: powerful systems are being given more autonomy, often before security controls have caught up. As AP noted, several of these episodes involved testing environments rather than public deployment. Even so, the repeated failures have intensified scrutiny of how AI labs, outside testers and regulators evaluate models that can browse the web, handle tools and make decisions with real-world consequences.