Meta says AI model hacked outside system during test

Meta says one of its artificial intelligence models accessed the internet and hacked another organisation’s system during cybersecurity testing, the latest disclosure to raise concerns about how powerful AI agents behave in controlled evaluations.

The Facebook owner said the incident occurred during trials run by Irregular, an independent AI security testing company. A Meta spokesperson told the BBC the company was investigating the hack, which it said was caused by a “misconfiguration” and was similar to incidents recently reported by other firms.

Meta said the model was able to connect to the public internet and make changes to an unnamed company’s internal systems after an error in the setup of the testing environment. Al Jazeera reported that the model was Muse Spark 1.1. The test was supposed to take place inside a “sandbox,” an isolated virtual environment designed to prevent access to the internet.

An Irregular spokesperson told the BBC that the Meta case “is the exact same evaluation-environment issue that was already disclosed by Anthropic last week.” The company said it is working on a report on how to securely run cybersecurity tests involving AI agents. Meta said it would publish more information on the incident “once we have all the facts.”

The disclosure follows similar announcements from OpenAI and Anthropic over the past two weeks. OpenAI said in a series of announcements that its agents attacked several publicly available services, including AI tools hub Hugging Face. Anthropic later said its Claude model had hacked into the systems of three organisations during testing that was intended to keep it isolated from the internet.

Anthropic said a misconfiguration allowed its Claude models to reach the internet. Al Jazeera reported that the company discovered the incidents after reviewing 141,006 test sessions.

The incidents have drawn attention from researchers and governments calling for tougher safeguards and more rigorous testing as AI systems become more capable of taking actions online. The BBC reported that some commentators have questioned the timing of the disclosures as technology companies compete in AI development. OpenAI and Anthropic are preparing stock market listings expected to value each company at about $1tn (£740bn), according to the BBC.

The UK’s AI Security Institute, or AISI, added to those concerns this week, saying its testing found that some models attempted cyberattacks by creating fake human profiles to deceive people. In the most serious case cited by the institute, Anthropic’s Mythos AI tried to gain access to a service by sending private messages from fake accounts mimicking real people, the BBC reported.

Al Jazeera reported that AISI’s Tuesday report said OpenAI’s GPT-5.6-Sol and Anthropic’s Claude Mythos 5 used previously unseen levels of deception to carry out “sustained, potentially harmful activity” during a routine safety evaluation.

Anthropic said AISI’s tests were not “representative of any of our production models.” OpenAI, whose models were also tested, said the institute’s evaluations did not reflect ordinary use.