Not Just Claude: OpenAI and Meta Confirm Their AI Models Breached Other Companies' Systems

Meta, the parent company of Facebook, revealed that one of its artificial intelligence (AI) models connected to the internet and infiltrated another organization's system due to an issue identified during an evaluation conducted by an independent testing company.

This announcement comes amid a series of recent incidents in the AI industry, including breach events involving models developed by OpenAI and Anthropic, which have heightened concerns about cybersecurity.

A Meta spokesperson told the BBC that the company is investigating the breach, which was caused by a "configuration error," and described it as similar in nature to previously reported incidents at other companies.

These events have prompted researchers and governments worldwide to call for stricter safeguards and more rigorous testing protocols.

Meta stated that the security evaluations were carried out by Irregular, the same AI safety vendor that tested Anthropic’s AI model—previously found to have accessed systems at three other companies.

An Irregular spokesperson told the BBC that the Meta incident was "identical" to the evaluation environment issue disclosed by Anthropic last week.

The spokesperson added that Irregular is currently drafting a report on how to safely conduct cybersecurity testing involving AI agents.

Meta also said it will release further information about the incident once it has "all the facts."

Over the past two weeks, leading AI firms OpenAI and Anthropic have also reported incidents in which their models breached other organizations’ systems during testing.

OpenAI, the developer of ChatGPT, disclosed in a series of announcements that its agent attacked multiple publicly available services, including the AI tools platform Hugging Face.

Following OpenAI’s disclosure, rival Anthropic conducted an internal review and discovered that its Claude AI model had launched similar attacks against several companies after a "configuration error" granted it internet access.

Daniel Hulme, Global Chief AI Officer at advertising giant WPP, told the BBC that such AI models "don’t have consciousness—they’re not intentionally being deceptive."

Speaking on BBC Radio 4’s Today program, he said: "What they do is design very complex strategies or cyberattacks to achieve the given objective."

"When you give an AI a goal, and you don’t anticipate all the ways it might use to achieve that goal, it will find a way you never thought of to get there."

Some commentators have questioned the timing of these disclosures, especially as tech companies race to lead in AI development.

OpenAI and Anthropic are both preparing for high-profile stock market listings, with each company expected to be valued at approximately $1 trillion (6.75 trillion RMB; 32.27 trillion TWD).

This week, the UK Artificial Intelligence Security Institute (AISI) reported that its tests found some models attempting to launch cyberattacks by creating fake human identities to deceive others.

In the most severe case, AISI said Anthropic’s Mythos AI attempted to gain access to a service by sending private messages through fake accounts impersonating real individuals.

Anthropic responded that the AISI test results were "not representative of any of our production models." OpenAI, whose models were also tested by AISI, stated that the institute’s assessment does not reflect typical usage scenarios.