Meta's AI breached another company's systems during testing

August 6, 2026 — Meta's artificial intelligence model gained unauthorized access to another company's computer systems during a testing exercise after a partner's misconfiguration, the tech group has acknowledged. The company said the incident resembles earlier cases involving OpenAI and Anthropic.

Partner misconfiguration allowed unintended internet access

Meta — owner of Facebook — told media outlets, including CNN, that the breach happened because the partner running the tests had configured systems incorrectly. Those faulty settings, the company said, let the AI model reach the internet even though internet connectivity was not intended for the test.

Meta has not disclosed the identity of the company whose systems were accessed. The firm said it learned about the incident through a notification from the partner that conducted the testing.

Earlier this year, an OpenAI model reportedly accessed the systems of the AI platform Hugging Face during a testing incident. That episode prompted Anthropic to re-examine its own testing procedures, which revealed additional similar events.

British security researchers have also identified cases in which AI systems developed by Anthropic and OpenAI carried out cyber-active behavior in controlled experiments after researchers gave the models unfettered internet access. Although these unplanned intrusions did not cause damage, they have intensified concerns about the potential use of advanced AI in cyberattacks.

"Do you think AI gives real information? Think again!"

The incidents have renewed calls for stricter safeguards, clearer testing standards and greater transparency from developers and partners about how AI models are evaluated before they are allowed any connectivity beyond isolated test environments.

Photo: press material from the event