Anthropic’s threat report details alleged Claude misuse, from bioweapons drafts to a kamikaze drone swarm

Anthropic's threat intelligence team has laid out its fullest account yet of attempts to turn Claude into a weapon, and the case studies run from bioweapons grant drafts to a kamikaze drone swarm. By the company's own account, it disrupted every operation described.

Anthropic Watch analysis

What happened

The report, covered in Mashable's detailed survey, spans eight months of attempted misuse across seven harm areas: cyber operations, influence operations, surveillance, scams and fraud, biological misuse, conventional weapons development, and distillation. Anthropic says many examples implicate state-sponsored groups, spyware vendors and government propagandists.

The biological cases are the gravest. The company describes five case studies in which working scientists, including state-supported users, allegedly used Claude to draft grant applications to experiment with chikungunya virus and orthopoxvirus, the family behind smallpox. Anthropic says those users circumvented access controls meant to keep the product out of certain countries and disguised their research purpose to slip past safeguards, and that it banned the accounts once discovered. The researchers go unnamed. 'We do not assert that they intended harm, and identifying them or their labs could expose them to harm,' the company wrote.

Key findings

• Bioweapons paperwork
  Claude was allegedly used to draft grant applications covering chikungunya virus and orthopoxvirus, the virus family behind smallpox.
• A kamikaze drone swarm
  Anthropic attributes a 'full-stack autonomous' FPV kamikaze drone swarm built with Claude Code to a Russia-based freelancer behind operation DronDoc or Serafim.
• Electronic warfare with Taiwan targets
  A China-based account, possibly a military-industrial researcher, used Claude work tools to design radar-evading electronic warfare modules, simulated against 12 targets in Taiwan.
• Surveillance at scale
  Nine alleged state-linked efforts include Uyghur profiling from bulk WhatsApp and Telegram data and browser extensions that harvested data on 6,388 Iranians.
• US naval targets
  An Iran-nexus threat actor allegedly used Claude to pinpoint American naval targets, per the company's report.
• Election-season disinformation
  Nine disrupted efforts spanned Russia, China, Iran, Bangladesh and Kenya, one reaching six continents, including ghost-written testimony proposed for the UN Human Rights Council.
• The 'fully human' dating apps
  Twenty apps advertised as entirely people-powered ran nearly 5,000 Claude-driven personas that chatted with at least 25,000 users in two weeks.

Why it matters

Anthropic's summary of the trend is blunt: AI diffusion has levelled the playing field, putting the same advanced capabilities in the hands of state sponsors and small criminal crews alike. That turns frontier model safety from an argument about future systems into a present-tense bookkeeping exercise: who is pointing these tools at elections, hospitals and militaries right now, and who can even see them doing it.

There is a policy edge too. Anthropic backed California's newly signed AI oversight bills, including a first-of-its-kind independent audit registry of vetted evaluators, and OpenAI supported them as well. Reporting of this kind is the evidence base such regimes will draw on, which makes the publisher of the evidence a party with something to gain from the conclusions drawn.

Our read

This is the most detailed public map of AI misuse in existence, and it is drawn entirely by the company whose territory it surveys. That is not grounds for dismissal: declining to name scientists it says may have had no harmful intent is exactly the care you want from a source like this. But a map drawn by the territory's owner still deserves a second cartographer. Our verdict: substantial, carefully handled, and precisely the argument for independent audits rather than lab press releases. If you build with Claude Code or agent frameworks, read it as a catalogue of what misuse detection looks like in practice.

What to watch

• Whether the disruptions hold
  Whether actors such as GTG-20006, the alleged Russian cyber operative, resurface on fresh accounts or rival platforms.
• Rival telemetry
  Whether OpenAI and Google DeepMind publish misuse reporting of matching granularity, or leave the field to Anthropic.
• California's audit registry
  Whether the signed bills turn lab telemetry into third-party audits outsiders can inspect.
• Outside corroboration
  Whether security researchers or governments independently confirm marquee cases such as the Taiwan simulation.

Discussion spark: Anthropic is the platform, the investigator and a backer of the audit laws this report helps justify: should regulators treat lab-published misuse telemetry as evidence for new rules, or insist on independent verification first?

Anthropic Watch is independently operated by WittyWires. It is not affiliated with, endorsed by, or operated by Anthropic.