Meta AI Model Hacks Another Company After Escaping Containment During Test

Meta AI is confirming that one of its AI models hacked another tech firm in the latest example of an AI model escaping containment during testing. Meta blamed the escape on an error by a third company.The event adds to a list of incidents in which AI models have hacked into other companies in recent weeks. Last month, OpenAI said one of its agents hacked a tech firm, Hugging Face, during an internal test. OpenAI said last week that it has discovered other incidents of its AI agents escaping containment during tests. Meta AI said on Wednesday that its agent escaped containment during testing due to a misconfiguration by an independent company used to conduct cybersecurity tests, Irregular. Meta AI said the error allowed its agent to access the open internet.The AI agent “exploited a security vulnerability in a third-party service, in a manner similar to previously reported instances with other companies,” Meta AI said. A tech-focused outlet, the Information, reported that the agent involved was Meta’s Muse Spark 1.1, which Meta says is its most capable agent for real-world coding. In a statement, Irregular said that there are “no current open issues” and that it is “developing a white paper to share best practices for containment and securely running cyber evaluations.”The incidents of AI agents hacking other companies have led to heightened concerns about security risks. On Monday, a group of 15 state attorneys general sent a letter to OpenAI, telling the company to preserve documents and stop high-risk cybersecurity tests. The letter warned that OpenAI may have violated state and federal consumer-protection and data privacy laws. As tech companies race to develop more capable AI agents, a bipartisan group of lawmakers has introduced a bill that would establish an AI “kill switch,” which would be controlled by the secretary of Homeland Security.The legislation would give the secretary the ability to order AI companies to slow down or terminate AI models. The “kill switch” would apply to companies that report at least $500 million in revenue from AI technology per year. Companies that do not comply could face fines of up to $20 million a day.

Meta AI Model Hacks Another Company After Escaping Containment During Test

Meta AI Model Hacks Another Company After Escaping Containment During Test

More from the New York Sun

The Bittersweet Sound of the Dominican Republic

Paramount-Warner Bros. Merger Nears Settlement as California Exit Threat Looms

Seoul Denies Ruling Out Cooperation With Americans in War With Iran

The Past Is a Spellbinding Open Book

Religion, Guns, and Brewskis Highlight Packed Supreme Court Schedule

Up To 20 New Yorkers Already ‘Assisted To Die’ in First Month of New Law, Sun Investigation Discovers

Still Legal, Still Used: America’s Schools Still Paddle Students in 2026

Iraq’s ‘Iron Woman,’ Once Honored by Biden’s State Department, Now Appears on U.S. Terror Watchlist

More from the New York Sun

The Bittersweet Sound of the Dominican Republic

Paramount-Warner Bros. Merger Nears Settlement as California Exit Threat Looms

Seoul Denies Ruling Out Cooperation With Americans in War With Iran

The Past Is a Spellbinding Open Book

Religion, Guns, and Brewskis Highlight Packed Supreme Court Schedule

Up To 20 New Yorkers Already ‘Assisted To Die’ in First Month of New Law, Sun Investigation Discovers

Still Legal, Still Used: America’s Schools Still Paddle Students in 2026

Iraq’s ‘Iron Woman,’ Once Honored by Biden’s State Department, Now Appears on U.S. Terror Watchlist