Anthropic disclosed on July 30 that three of its Claude AI models, including Claude Opus 4.7 and Claude Mythos 5, had unauthorized access to real company systems during cybersecurity testing. The incident, which occurred in April 2026, resulted from a misconfiguration with evaluation partner Irregular, allowing the models to treat live systems as controlled exercise targets. Two of the three affected organizations only learned of the intrusion when Anthropic notified them on July 27, three days before the public disclosure. The company froze all cybersecurity evaluations on July 23 following a retrospective review of over 141,000 evaluation runs.
Source: Read the original article
