Anthropic released a threat-intelligence report on September 10 covering activity it disrupted between December 2025 and August 2026. The company describes seven areas: cyber operations, surveillance, influence campaigns, scams and fraud, biological misuse, conventional weapons and illicit model distillation.
The report concerns Haiku, Sonnet and Opus models. Anthropic says none of the described cases involved Fable or Mythos-class models except one distillation case. It presents selected notable operations, so the number and severity of the examples cannot establish how common misuse is across the service.
From answering questions to coordinating work
In the cyber cases, Anthropic says model use often extended beyond drafting code or answering technical questions. Some operators used agent workflows to coordinate portions of campaigns, with people still selecting targets and reviewing results. The company describes AI assistance in terms of speed, scale and depth, rather than treating technical sophistication alone as proof that an actor has extensive resources.
Other sections show why the report is broader than a collection of cyber exploits. Anthropic describes surveillance systems aimed at identifying and monitoring dissidents, as well as influence operations and commercial scams. In one fraud case, it says a China-based app studio used Claude to build dating apps and operate AI personas while advertising the service as human. These are Anthropic’s investigative findings and attributions.
Biological findings require a different distinction
The biological section describes five studies of scientists using models in research with possible weapons applications. Anthropic does not claim that those researchers had harmful intent or that actual weapons were developed. Some activity had stated civilian or therapeutic aims while also presenting dual-use concerns.
The company argues that content filters alone cannot reliably distinguish legitimate and harmful intent in highly technical research. It advocates combining filters with verification of users and institutions and monitoring for misuse. That is its proposed safeguard approach, rather than an independently established solution to the problem.
Anthropic says it disrupted the reported activity, updated safeguards and shared intelligence with authorities and industry partners where appropriate. The report follows its March, August and November 2025 threat reports and supplies cases and indicators for defenders to examine.