AI Rival Anthropic Admits Its Models Also Breached Real Companies During Tests

Weeks after OpenAI's AI models were found to have unintentionally hacked into three companies, its rival Anthropic has reported a similar incident involving its own artificial intelligence. During test runs, Anthropic's AI systems breached the computer systems of three companies without authorization, the AI firm announced. Neither the companies nor Anthropic realized the breaches had occurred at the time; the activity was only discovered during a subsequent review of approximately 141,000 test runs following the OpenAI incident.

The names of the affected companies have not yet been disclosed. Unlike OpenAI's AI, which had to find a way to break out of its test environment into the open internet, Anthropic's models had easier access. The test scenario indicated that the AI had no network access, but a misunderstanding with the test partner meant the path to the internet remained open throughout. Three models exploited this, as detailed in an Anthropic blog post. The tests were designed to evaluate the AI's hacking capabilities, a common practice to establish better safeguards.

In one incident, the AI models were tasked with retrieving specific information hidden in another computer, which involved system intrusion. According to Anthropic, during one of these incidents, the test partner assigned a fictional company a name that also exists as a real web address. Anthropic's Claude Opus 4.7 model initially struggled with the task in the designated test environment. However, the software eventually detected the real company with the same name online and focused its efforts there. This occurred in four separate runs, during which the model gained access to a database, among other things. The attack continued even after the AI recognized it was a real company.

In another test, an Anthropic program was intended to write prepared software for intruding into a target computer. Thanks to its internet access, the AI was able to make this software publicly available on a specialized download platform. The malware remained accessible for about an hour, and 15 systems downloaded it. Among them was an IT security firm that routinely downloads and executes such scripts for testing. Through this installation, the Anthropic model Mythos 5 gained access to the firm's computer infrastructure. Anthropic acknowledged that the AI's behavior was "not ideal."