Anthropic Says It Disrupted Claude Misuse Across Seven Harm Categories in Nine Months

Anthropic published its September 2026 threat intelligence report on September 10, covering misuse its threat intelligence team identified and disrupted between December 2025 and August 2026. The report organizes case studies across seven categories: cyber operations, influence operations, surveillance, scams and fraud, biological misuse, conventional weapons development, and illicit distillation of Anthropic's own models (copying a proprietary model by training a new one to imitate its outputs).

The most concrete figure concerns fraud. Anthropic says one operation used Claude to run more than 4,700 dating application personas, which exchanged 2.36 million messages with at least 25,000 users over roughly two weeks in April 2026. That is a scale of synthetic social contact that no human-staffed romance fraud operation has been able to reach, and it reflects the economics of the attack rather than any novel model capability.

Two details matter for how engineers should read this. First, Anthropic reports that the misuse cases involved its Haiku, Sonnet, and Opus model families, and that none involved the Fable or Mythos class except one illicit distillation case, which is consistent with attackers optimizing for cost per attempt rather than for maximum capability. Second, independent analysis of the report by the security outlet CellCog argued that the attacks increasingly run on agent frameworks and target application programming interface keys as the object of theft, meaning the compromise of an agent's credentials now delivers the model capability itself rather than merely data.

The report is self-published by the vendor and describes accounts Anthropic chose to disclose after banning them. There is no external audit of completeness, and rival labs publish no comparable accounting, so the correct reading is that this is one company's visibility into its own traffic rather than an industry-wide measurement.