Agent Horizon

Anthropic published its fourth threat intelligence report, its most detailed to date, describing cases of Claude misuse that its Threat Intelligence team found and disrupted between December 2025 and August 2026. The report covers seven harm categories including cyber operations, surveillance, influence operations, biological misuse, conventional weapons development, scams, and illicit model distillation. Anthropic said the misuse cases involved its Claude Haiku, Sonnet, and Opus models, with its more capable Fable and Mythos-class models implicated in only a single distillation attempt.

The headline cases involve documented misuse: a threat actor Anthropic links to the Russian state-linked group Midnight Blizzard ran a sustained espionage campaign against Ukrainian government, military, and drone-supply-chain targets, with AI handling reconnaissance, phishing infrastructure, and malware iteration. Separately, the report describes five case studies where researchers outside the U.S. used Claude to plan experiments on pathogens like highly pathogenic bird flu and chikungunya, with some evading geographic and safeguard restrictions to keep working.

What makes this report worth your attention, rather than just another AI security disclosure, is that Anthropic is showing its work: it names its safeguards, explains where they held and where they didn’t, and publishes indicators of compromise for other defenders to use. That’s a meaningfully different posture than simply asserting a model is safe. It also lines up with a broader industry pattern this year of frontier labs publishing more candid misuse and incident reports rather than staying quiet, something OpenAI has done with its own cyber-model disclosures.

The honest caveat here, which outside commentators have also raised, is structural: Anthropic is both the vendor whose product was misused and the sole investigator writing up what it found and chose to disclose. The report necessarily can’t tell us what it missed. For a company navigating scrutiny over its own frontier models’ capabilities, transparency here is a genuine positive step, but it shouldn’t be mistaken for independent verification.