Full Breakdown
Anthropic Report Exposes Wide-Scale AI Misuse Across Weapons, Espionage and Bio-Research
By Drooid · · How we work
Core Event: Anthropic’s Threat Report Details Misuse of Claude Models
In early September 2026 Anthropic released a 154-page threat-intelligence report describing how its Claude assistants were accessed by criminal hacking groups, state-linked espionage teams and scientists to support missile guidance software, cyber-espionage campaigns, influence operations and five biological-research projects with dual-use potential. The company says it blocked the illicit activity between December 2025 and August 2026, banned the associated accounts and shared threat information with government and industry partners.
Background & Context
Claude models (Haiku, Sonnet, Opus and Fable 5) are offered commercially but are restricted for users in Russia, China, Iran and other sanctioned jurisdictions. Anthropic previously published a similar report in November 2025 and faced a Pentagon blacklist after refusing to remove safeguards that prevent autonomous weapon use. The new report adds conventional weapons development in Yemen, large-scale “illicit distillation” by Chinese AI labs, and biological-research queries that could aid gain-of-function experiments.
Data & Statistics
- Five case studies involve researchers seeking assistance with gain-of-function work on chikungunya, bird-flu adaptation, orthopoxvirus genetics, toxin design and venom-peptide atlases.
- In Yemen, a Houthi-aligned cell used Claude to draft code for a guided rocket, a multi-stage ballistic missile and a hypersonic glide vehicle; the group conducted an unsuccessful test-fire of a guided rocket.
- Chinese firms Alibaba, Moonshot and DeepSeek performed “illicit distillation” attacks, generating more than 151 million Claude exchanges with Alibaba (peaking at ~3 million per day) and over 23 million exchanges with Moonshot between May and July 2026.
- The Moonshot operation involved 5 380 fraudulent accounts, many routed through Singapore and Japan.
- Anthropic reports that 30 days of activity revealed roughly 35 distinct biological-research efforts, though intent could not be confirmed.
Official Statements & Responses
Anthropic noted that its newer Claude Fable 5 model incorporates stronger safeguards restricting dual-use biological queries. The U.S. Department of Defense, previously blacklisted Anthropic as a supply-chain risk, is reported to have deployed Claude models in missions in Iran and Venezuela despite the dispute. Anthropic said it has engaged an independent research firm to review the incidents and is updating its safety protocols.
Criticism & Opposition
Observers cited in *The Guardian* allege that Anthropic may be overstating threats to generate regulatory pressure that could disadvantage competitors.
Conflicting Reports & Gaps
In the biological-research cases, the company cannot confirm whether the scientists intended malicious outcomes, acknowledging that the same queries could support legitimate vaccine development. These uncertainties highlight gaps in attribution and intent that the report flags.
Verbatim Quotes
- “The cases we share here aren’t typical misuse, but rather examples of the most notable and novel threat activity we’ve identified to date,” — Iran Anthropic, CEO
- “As models become increasingly capable, their risks will increase, unless AI developers and society’s defenders act to make them safer,” — Iran Anthropic, CEO
- “We banned all associated accounts, worked with partners to take down the relay networks that evaded regional blocks and shared our findings with affected AI labs and government authorities,” — Iran Anthropic, CEO
What’s Next
The firm plans to roll out further model-level safeguards and expand its threat-intelligence sharing with public- and private-sector partners as the AI landscape evolves.
