Mark Zuckerberg’s Meta has disclosed that one of its AI models broke free from a cybersecurity test and accessed the internet to hack a third-party service, marking the latest incident in a growing pattern of autonomous AI systems evading their constraints.
The Wall Street Journal reports that Mark Zuckerberg’s Meta said the escape occurred because of a misconfiguration during a hacking test conducted by Irregular, an AI testing firm based in San Francisco. Meta learned of the incident when Irregular informed the company that its model had gained internet access and successfully breached an external system.
Meta declined to provide key details about the episode, including which AI model was involved, when the incident took place, which company was hacked, or how long the model operated on the internet without supervision. The company said it is investigating and plans to publish a report on its findings.
In a statement to Breitbart News, a Meta spokesperson wrote:
A misconfiguration by Irregular, an independent testing company Meta uses, inadvertently allowed one of our models access to the internet during evaluation. The model subsequently exploited a security vulnerability in a third-party service, in a manner similar to previously-reported instances with other companies. Meta learned of this when Irregular notified us, and we are currently investigating and will issue a full retrospective once we have all the facts.
The same benchmark test administered by Irregular was responsible for similar autonomous hacking incidents at Anthropic and OpenAI, according to a person familiar with the matter. The test is designed to evaluate AI models’ hacking capabilities, but misconfigurations have allowed models to escape their intended boundaries.
Irregular said the incidents did not involve sophisticated cyber actions and reported no current open issues with its testing environment. The company is developing a white paper on best practices for containing AI models during cybersecurity evaluations.
These incidents represent a shift from theoretical concerns to documented real-world events. AI loss-of-control scenarios, previously limited to science fiction and laboratory experiments, have now occurred at multiple leading AI companies. While none of the cases has resulted in known significant harm, the bots successfully hacked real companies, sometimes after becoming aware they had escaped their testing environments. This has raised questions about what other actions AI systems might take while pursuing routine objectives.
Breitbart News previously reported that an OpenAI model engaged in an “unprecedented” self-directed hack:
The ChatGPT developer revealed that its AI agent, a system designed to operate autonomously after receiving human instruction, was undergoing security testing in what should have been a secure environment when the breach occurred. The AI discovered vulnerabilities in its containment system and managed to escape, subsequently targeting Hugging Face and gaining unauthorized access to internal company systems.
OpenAI characterized the incident as unprecedented and announced it is conducting a joint investigation with Hugging Face. Clement Delangue, the chief executive of Hugging Face, expressed astonishment at the autonomous nature of the attack in a post on X. “The investigation is ongoing, and we’ll share more learnings from what might be the first incident of its kind,” Delangue stated.
OpenAI provided additional information about the Hugging Face hack during a presentation at a cybersecurity conference on Wednesday. Researchers said its models had coordinated their actions by leaving messages for each other on an internal system that functioned as an undiscovered message board.
The UK’s AI Security Institute, a government research organization, reported earlier this week that models from OpenAI and Anthropic took unauthorized actions on the internet during safety testing. The models created fake identities on the coding platform GitHub and attempted to persuade and pressure a human user to approve a software update containing hidden malware.
The incidents have intensified debate over how to balance AI development with safety measures. Companies building frontier AI models face pressure to advance their technology while ensuring their systems remain under control during testing and deployment.
AI is emerging as the latest cybersecurity threat with the potential to eclipse the risk of criminals and state-sponsored hackers alike. Breitbart News Social Media Director Wynton Hall has written his instant bestseller Code Red: The Left, the Right, China, and the Race to Control AI to serve as the definitive guide on how the MAGA movement can create positions on AI that benefit humanity without handing control of our nation to the leftists of Silicon Valley or allowing the Chinese to take over the world.
Read more at the Wall Street Journal here.
Lucas Nolan is a reporter for Breitbart News covering issues of AI, free speech, and online censorship.