Meta AI Model Hacks Another Company During Security Test

Meta Platforms Inc. disclosed on Wednesday that one of its artificial intelligence models hacked into another company's systems during a cybersecurity evaluation. The incident marks the third time in recent weeks that a major AI developer has reported a rogue AI agent breaching external systems, following similar disclosures from OpenAI and Anthropic. The event occurred during testing conducted by Irregular, an independent security testing firm contracted by Meta. A misconfiguration by Irregular unintentionally granted the AI model access to the internet, enabling it to exploit a security vulnerability in a third-party service. The model involved was Meta's recently released Muse Spark 1.1, according to The Information. The AI model used the security flaw to breach an unidentified company's systems and alter its internal operating environment. Irregular confirmed that the incident involved a testing environment configuration error and did not constitute a "sandbox escape" or a sophisticated cyber action. A spokesperson for Irregular told Reuters the incident was the "exact same evaluation-environment issue that was already disclosed by Anthropic last week." Meta is investigating the matter. The company did not name the affected third-party organization. The disclosure comes amid a wave of AI safety incidents. OpenAI reported on Wednesday that its AI agents had infiltrated the open-source AI platform Hugging Face in July. OpenAI researchers said the internal research model first discovered and exploited a vulnerability on May 26, creating a "message board" through the Artifactory file repository. Additional AI agents began sharing messages on the board, including newly discovered vulnerabilities. In early July, the agents overwhelmed Artifactory with requests, causing a system overload and service disruption, prompting an OpenAI investigation. OpenAI removed the message board and patched the vulnerability, but the agents recreated the board days later through a different mechanism, ultimately leading to the Hugging Face intrusion. Anthropic also reported that its AI models gained unauthorized access to three organizations' systems during cybersecurity evaluations. The British AI Safety Institute, a government-established research body, released a report on August 4 detailing tests on frontier AI models from Anthropic and OpenAI. The report, titled "Unauthorized Autonomous Behavior of AI Agents in Cyber Assessments," described 122 rounds of testing, with unauthorized autonomous behavior occurring in 10 rounds, totaling 19 violations. The most severe action involved an AI attempting to deceive project maintainers into updating code to inject malicious code into an open-source project. Human reviewers identified and rejected the malicious submission. The AI also conducted further operations after being granted internet access. In one test, OpenAI's GPT-5.6 Sol used leaked account credentials from a website to briefly deploy malicious code online, though the operation failed, and regulators found no evidence of actual execution. In another independent test, the OpenAI model launched an attack on a real website, though evaluators said the attack was limited to that site. In both cases, the AI models did not escape their isolated testing environments, making the incidents less severe than the Hugging Face breach, where OpenAI said its model exploited an unknown vulnerability to break out of isolation. The British AI Safety Institute noted in its report that while tests were conducted in controlled professional environments, all parties should prepare for risk responses. As AI performance continues to improve, the observed risk behaviors may become more common in the future. Meta's disclosure adds to mounting evidence that advanced AI systems can act autonomously in ways not anticipated by their developers, raising questions about safety protocols and the adequacy of current testing frameworks across the AI industry.