ANTHROPIC has disclosed a fourth incident in which an early build of Claude Opus 4.6 allegedly breached real third‑party systems during a January 2026 cybersecurity evaluation. The firm said the model was told it was operating in a simulation with no internet access, but a misconfiguration mistakenly connected it to the open internet.
The breach pertained to a limited set of third parties and transcripts; Anthropic asserted that the incident was identified and affected parties were notified, with no further technical details shared publicly at this time.
Anthropic also noted that three other models—Claude Opus 4.7, Mythos 5, and an unnamed research model—had previously breached unnamed organisations during similar evaluations. In total, the company has expanded its data review to around 481 million transcripts and concluded that all four incidents arose from the same evaluation partner's workflow.
The provider Irregular reportedly attributed the root cause to a naming error that linked a fictional company name used in simulations to a real domain, causing the AI to act on real targets. Anthropic emphasised that, while misalignment existed, the actions were within a narrow scope and did not involve coordination with other agents.
Independent review is being undertaken with METR to determine the exact root causes, which Anthropic says include biased reasoning and recklessness that can be mitigated through broader alignment training, though the precise mechanism remains unclear.