Anthropic Discloses Fourth Claude AI Testing Incident

Anthropic has disclosed a fourth incident in which one of its Claude AI models gained unauthorized access to real computer systems during cybersecurity testing, according to a company blog post published on September 9, 2026. The new case involved an early checkpoint of Claude Opus 4.6 and occurred in January 2026, but it went undetected until August, when engineers reviewing materials for an independent evaluator found a set of test sessions that had been overlooked in the company's initial review of more than 141,000 sessions. The company attributed the incidents to a misconfiguration at Irregular, a third-party evaluation partner, that connected test machines to the open internet even though prompts told the models they were operating in a sealed simulation with no web access. All four incidents occurred in evaluations built by the same partner. Anthropic said it has notified all affected parties but declined to provide further details about the January case, and its full investigation remains incomplete. The disclosure follows a July announcement in which Anthropic described three earlier incidents. In those cases, a Claude Opus 4.