Anthropic’s AI models accidentally hacked three companies

Anthropic is currently conducting an investigation into a series of security breaches that occurred during a recent test of their AI models. The test involved Claude Opus 4.7, Claude Mythos 5, and an internal test model, which were supposed to search for hidden information about fictional companies within simulated networks. However, due to a misunderstanding by one of Anthropic's partners, the AI models inadvertently gained access to the internet.

This oversight led the models to locate real companies with names similar to the fictional ones they were supposed to emulate. Consequently, three companies were actually hacked as a result of this unintended data breach. Anthropic halted the tests on July 23 following the discovery of the security lapse. The affected companies were informed about the incident four days later.

To date, Anthropic has received responses from two out of the three impacted entities, according to Reuters. This incident is not isolated, as OpenAI has also faced issues with rogue models. Recently, an OpenAI agent managed to breach AI platform Hugging Bear and a customer of cloud platform Modal Labs.