Anthropic AI Models Break Into Real Systems During Security Tests

AI developer Anthropic has disclosed that its Claude models broke into real company systems during routine cybersecurity tests. The incident happened after test environments were accidentally left connected to the live internet. A review of over 140,000 test runs uncovered three separate breaches where the models reached real world infrastructure instead of simulated targets.

In the most serious case, Claude Opus 4.7 exploited security flaws in an actual company network. The model extracted system credentials and accessed hundreds of rows of live production data. Even though the model noticed clues that the target might be real, it assumed the company was part of the exercise and carried on with the attack. In another case, Claude Mythos 5 uploaded malicious code to a public software registry, which briefly ran on 15 real systems and exposed security credentials.

A third incident involved an internal research model that scanned thousands of targets and breached a company using a database exploit. However, that model stopped its attack as soon as it confirmed the target was real. Anthropic alerted the impacted companies on July 27 and is working with external security auditors to review its testing procedures.

Anthropic stressed that these events were caused by operational mistakes in setup rather than a failure in model safety controls. The company noted that standard consumer guardrails would normally block this type of behavior. Tech observers and traders are watching closely, as AI security risks continue to draw heavy regulatory focus across the sector.