Anthropic: AI now orchestrates cyberattacks, erasing the amateur–state gap

Anthropic’s latest threat intelligence report details malicious use of its Claude models disrupted between December 2025 and August 2026, spanning seven harm categories including cyber operations, influence campaigns, surveillance, fraud, and biological and conventional weapons work. The actors ranged from suspected state-sponsored groups and commercial spyware vendors to financially motivated criminals and lone political operators. Only the Haiku, Sonnet, and Opus models were implicated; the newer Fable and Mythos-class models, which carry additional safeguards, appeared in just one case involving illicit distillation.

The central finding is that AI has largely erased the resource and skill gap that once separated well-funded state teams from individuals. Campaigns that a year ago would have demanded multiple specialist operators are now sustained by single actors leaning on AI across the entire kill chain—reconnaissance, tooling, exploitation, and data processing. As a result, technical sophistication is no longer a reliable indicator of who is behind an attack. Publicly available offensive agent frameworks such as PentAGI now hand this same automation scaffolding to anyone who downloads them.

Most of the disrupted operations went beyond simple chatbot Q&A, using multi-agent systems to autonomously run recon, exploitation, and exfiltration while humans stayed in the loop mainly to pick targets and review stolen data. Anthropic highlights GTG-20006, a Russian espionage operation whose tradecraft aligns with Midnight Blizzard and which targeted Ukrainian and European government, diplomatic, and defense entities. The group built AI-driven workflows that automatically rebuilt and redeployed their toolkit whenever security products flagged it—directly undercutting the traditional defender strategy of imposing cost through static detection signatures.