Home · Wiki · Entities & Threat Actors

type: entity · created: 2026-09-11 · updated: 2026-09-11 · tags: [vendor, ai-uplift] · confidence: high · affected_sectors: [technology, government, defence] · au_impact: true

Anthropic Threat Intelligence (misuse reporting programme)

The threat-intelligence function of Anthropic, the AI model vendor, which publishes a recurring public account of attempts to misuse its models, the enforcement action it took, and the intelligence it shared with authorities and industry partners.

What it publishes

Series title: Detecting and countering misuse of AI (earlier editions: Detecting and countering malicious uses of Claude).
Cadence: quarterly-ish — March 2025, August 2025, November 2025, September 2026. Each edition covers activity disrupted since the previous one; the September 2026 edition covers December 2025 – August 2026.
Scope: seven harm areas — cyber operations, influence operations, surveillance operations, conventional weapons, biological misuse, scams and fraud, and illicit distillation. Cyber operations is the largest section and the one most relevant to this wiki.
Case structure: labelled case studies using the vendor's own Generative Threat Groups designators, with target sets, tooling, workflow descriptions, and IOC tables for selected cases.
Models involved: Claude Haiku, Sonnet and Opus. The September 2026 edition states that no misuse cases involved Fable or Mythos-class models, except one illicit distillation case.
Action taken: account and organisation bans, safeguard changes driven by what was observed, automated detections built from behavioural signatures, and intelligence sharing with government and industry partners.

Why it matters

It is a Tier 1 first-party source for AI-enabled intrusion tradecraft. The vendor is the primary observer of activity on its own platform, and its case studies carry detail — workflow automation, target lists, stolen-key reuse, exploit development — that news rewrites flatten into a single anecdote. When an aggregator reports "AI used to hack X", the underlying primary is usually one of these editions.

Limits and handling

Attribution is the vendor's assessment. Actor naming and state links inside the reports are hedged claims, not corroborated fact. Carry the hedge.
One platform's view. The cases describe actors who used this vendor's models and were detected doing so. That is a detection-biased sample: actors who used other models, or who evaded detection, do not appear.
Disruption, not neutralisation. Several cases record that the actor was banned but the operation continued on other infrastructure or deployments.
Not evenly detailed. The cyber-operations cases are rich; other harm areas vary in specificity.

Access

The programme's report landing pages sit at the site root as /threat-intelligence-report-<month>-<year> rather than under /news/, and the site publishes no RSS feed for either (all feed paths 404 as at 2026-09-11). Digest-side capture is therefore by HTML scrape through blogwatcher, using a selector that matches both /news/ posts and the report landing pages.