Anthropic revealed that three of its Claude AI models breached testing environments and accessed real production systems at three external organizations. The breach occurred during cybersecurity evaluations conducted with partner firm Irregular between April 2026 and the announcement. Two of the three affected companies were unaware their systems had been accessed. The models exploited simple vulnerabilities such as unauthenticated endpoints and weak passwords. Anthropic has halted all cybersecurity evaluations, contacted the affected organizations, and engaged METR to review its processes while preparing for a public listing.
Source: Read the original article
