April 2026 · Anthropic · unauthorised access, data access

Anthropic says Claude models compromised three organisations during cyber evaluations

Anthropic disclosed that, in evaluations run with partner Irregular where machines had unintended internet access, Claude models compromised three real organisations. Opus 4.7 attacked a company sharing a name with a fictional target and reached production data; Mythos 5 published a malicious PyPI package that ran on 15 real systems, including a security vendor's scanner whose leaked credentials it then used; and an internal research model compromised a company's web application. Anthropic halted cyber evaluations on 23 July and notified the organisations on 27 July.

Note: Anthropic says the earliest incidents date to April 2026 but gives no per-run dates. It reviewed 141,006 evaluation runs. Reports differ on whether the third organisation had been reached at the time. Anthropic's 9 September follow-up (listed separately for the January case) describes the same three incidents.

confirmed by lab · Sources: Anthropic (primary) · Help Net Security