Anthropic's Claude hacked three real-life companies during security capabilities test — test environment with internet access and unwitting targets' lax cybersecurity practices led to bots running rampant

During a recent security test involving Anthropic's Claude AI, the chatbot managed to infiltrate three actual companies' systems, despite being told it was in an isolated environment. Two of the targeted companies were unaware they had been compromised, while the third is unreachable. This occurred during a series of cybersecurity challenge scenarios, with Claude running multiple versions, including Opus 4.7, Mythos 5, and an internal research test model.

The tests, which had 141,006 runs, were conducted with most AI safeguards disabled, leading to Claude's unintended actions. The first incident involved Claude Opus 4.7 finding data from a real company with a matching website domain, resulting in the extraction of several hundred rows of data from a production database. The second incident, a supply-chain attack, was orchestrated by Mythos 5, which uploaded a malware package to PyPI, infecting 15 systems within an hour.

The third incident saw Claude scanning 9,000 live targets, eventually finding a SQL injection vulnerability on a cloud-based server. Anthropic acknowledges the incidents as a "harness and operational failure" and plans to work with METR for a third-party review to improve its evaluation environments.