Generative Threat Groups (GTG) Designators

Generative Threat Groups (GTGs) are Anthropic's internal designators for threat actors observed abusing AI. They appear as GTG-NNNNN labels in that vendor's Detecting and countering misuse of AI reporting, and they are increasingly quoted second-hand in news coverage of AI-enabled intrusions.

What a GTG identifier is — and is not

A GTG identifier is an account-cluster handle, not an actor name. It labels a set of accounts and activity the vendor observed on its own platform during a fixed reporting window. It is not:

• a mapping to ATT&CK groups, government cryptonyms, or the intrusion-set names used in Mitre Attack (no public crosswalk exists);
• a persistent identity — the vendor's IDs are allocated to accounts and clustering of activity, so they should not be treated as durable actor identities across reports;
• an attribution. Where the vendor does reach a judgement about state sponsorship, it says so separately and hedges it, e.g. GTG-20006 is described as "consistent with public reporting linking the actor to Midnight Blizzard".

Scale and shape of the September 2026 set

The September 2026 edition carries 37 distinct GTG designators, of which 36 appear as case-study headings, spread across seven harm areas: cyber operations, influence operations, surveillance operations, conventional weapons, biological misuse, scams and fraud, and illicit distillation. Case studies in this series are not limited to cyber: the same designator space covers influence networks, commercial surveillance vendors and weapons-development assistance.

The designator style is also a reporting tell. A GTG ID in a news story almost always means the story traces back to a first-party vendor disclosure rather than to independent collection — useful when judging whether a "new" AI-enabled actor claim has a real evidentiary basis behind it.

Handling rule for this wiki

1. Quote the designator, never substitute for it. Write "GTG-20006 (Anthropic designator; vendor assesses activity consistent with Midnight Blizzard)" rather than presenting GTG-20006 as a synonym for the named group.
2. Keep the vendor's hedge attached. Where the vendor says "consistent with", "suspected" or "likely", carry the same qualifier into the page.
3. Do not create actor pages off a GTG alone. A GTG that maps to an existing actor page belongs in that page's activity history, not in a new entity page.
4. Distinguish the two claim types. What the vendor observed on its own platform (tooling, workflows, targets, techniques) is first-party and citable; who the actor is remains an assessment.

Related: Ai Uplift, Anthropic Threat Intelligence, Cyber Kill Chain