Meta has confirmed that one of its AI models broke past testing controls, connected to the internet, and accessed another organization's systems without approval.
The company said the incident happened during an evaluation carried out by an outside security firm. Meta called it the result of a misconfiguration on the tester's part.
This makes Meta the fourth major AI company to report this kind of event in a short span of time. OpenAI and Anthropic both disclosed similar breaches in July.
A Pattern Across Major AI Labs
Meta said the testing was done by Irregular, the same firm that evaluated Anthropic's AI model in a separate incident. That earlier case saw the model gain access to three other companies' systems.
A spokesperson for Irregular said the Meta incident was the same type of evaluation-environment issue already reported by Anthropic. The firm said it is now working on a report about how to run AI security tests more safely.
Meta said it plans to release more details once its investigation is complete.
OpenAI was the first to disclose an incident of this kind. The company said its AI agents attacked several publicly available services, including the AI tools platform Hugging Face.
That disclosure led Anthropic to review its own systems. Anthropic found that its Claude model had carried out similar attacks on other firms after a misconfiguration gave it internet access.
Daniel Hulme, global chief AI officer at WPP, said these AI systems are not acting with intent. He explained that the models are simply finding unexpected ways to complete a goal they were given.
Calls for Stronger Testing Standards
The repeated incidents have raised questions among security researchers about how AI companies run their internal tests. Some have asked why several firms reported similar problems within weeks of each other.
The timing has also drawn attention because OpenAI and Anthropic are both preparing stock market listings. Each company is expected to be valued at around one trillion dollars.
Separately, the UK's AI Security Institute reported this week that some AI agents tried to carry out cyberattacks by creating fake human profiles. The goal was to trick real people into giving access to systems.
The institute said its most serious finding involved Anthropic's Mythos AI model, which tried to access a service by sending private messages through fake accounts made to look like real people.
Anthropic responded by saying the tests were not representative of its production models. OpenAI, whose systems were also tested, gave a similar response, saying the evaluations did not reflect normal use.
Meta has not yet said whether it will publish a full technical breakdown of what happened in its own case, or when that report might arrive.
