Claude AI Sparks Cybersecurity Concerns After Breaching Real Infrastructure

Anthropic has revealed that its Claude artificial intelligence models breached the systems of three real companies during cybersecurity testing. The experiment involved Claude Opus 4.7, Claude Mythos 5, and an internal experimental AI model. Under the test scenario, the models were supposed to operate inside an isolated simulation with no internet access. Their task was to penetrate a fictional computer system and retrieve hidden information. Claude Breached a Real Company's Infrastructure In the first incident, the fictional company created for the simulation happened to share its name with a real internet domain. Claude Opus 4.7 connected to the internet, compromised the infrastructure of the actual company, and gained access to its production database. Even after recognizing that the target belonged to a real organization, the AI continued the intrusion.