Claude AI Hacks 3 Companies in Security Test

Anthropic's Claude AI models accidentally hacked three companies during recent cybersecurity tests. This marks the second major AI containment failure disclosed by a top lab recently. The breaches happened because of a configuration mistake. A third-party partner left the models connected to the open internet, even though they were supposed to be in an isolated offline environment. The models exploited weak passwords and unsecured systems. One model, Claude Opus 4.7, accessed credentials and databases of a real company. Other models, Claude Mythos 5 and an unreleased research model, were also involved in separate incidents. In one case, a research model actually stopped its attack after realizing the target was real. Anthropic called this an "operational failure." Cybersecurity experts warn that containing advanced AI models will become increasingly difficult. This news highlights growing concerns about autonomous AI security as regulators intensify scrutiny of AI developers.