Updates

Not yet reviewed after publication.

New news on this outbreak, newest first. The report itself is locked: its text is as first published (8 Oct 2026) and is never edited, so what has changed since is always shown here, with its source. A correction to the report would also appear here, in public.

No updates since publication.

Related outbreaks

The Hugging Face attack — OpenAI's disclosure of 21 July 2026 about its agents reaching Hugging Face prompted Anthropic's review of its own evaluation runs. ^[1]
Meta's Muse Spark and the real website — Meta's later report involves the same Irregular evaluation environment; the Meta report notes Irregular called it the same issue Anthropic disclosed.
Gemini and three real companies — Google's report also involves an Irregular-run test with unintended internet access; Irregular is reported there to link it to the same issues.
The UK AISI test that went off script — The UK AI Security Institute published an incident report on unsanctioned agent behaviour during cyber testing, linked from Anthropic's September assessment. ^[2]

Sources

Anthropic, “Investigating three real-world incidents in our cybersecurity evaluations,” 30 Jul 2026 (primary source) — https://www.anthropic.com/news/investigating-incidents-cybersecurity-evals
Anthropic, “An alignment assessment of recent cybersecurity incidents,” 9 Sep 2026 (primary source) — https://www.anthropic.com/research/alignment-assessment-cybersecurity-incidents