Anthropic published its fourth threat intelligence report, "Detecting and Countering Misuse of AI: September 2026," on September 10, documenting cases of attempted misuse of Claude spanning bioweapons-related research, state-linked espionage and autonomous weapons software collected between December 2025 and August 2026.

Bioweapons research and a drone-swarm operation

The report says Anthropic identified and blocked five separate attempts by scientists to use Claude for research that could contribute to biological weapons development, including one gain-of-function study reportedly intended for a military research institute. Anthropic said newer Claude models can no longer be assumed to fall below the threshold for providing meaningful bioweapons-relevant assistance, a shift from the company's earlier public position.

Separately, the report describes a small team of apparently freelance developers based in Russia who used Claude to help write software for a swarm of autonomous first-person-view "kamikaze" drones, including terminal guidance, target selection and coordination logic between multiple aircraft, with the drones intended to select their own targets and detonate on impact without a human final decision. Anthropic said the group repeatedly targeted a location in Ukraine's Donetsk region in its testing and used VPNs to route around the company's geographic restrictions before being detected and cut off.

A tracked espionage campaign

Anthropic also detailed a sustained espionage campaign it tracks as GTG-20006, whose tradecraft and targeting the company says are consistent with the publicly documented Russian state-linked actor known as Midnight Blizzard. According to the report, operators used AI at nearly every stage of the campaign — including phishing, hijacking hotel Wi-Fi networks and compromising WhatsApp accounts — against government, military and diplomatic targets in Ukraine and allied European states, and built a system that used AI to automatically detect when its malware was flagged by security tools and rewrite the code until it evaded detection. One operator was identified by the handle "JackPoterz."

Why it matters

The report is Anthropic's most detailed public accounting yet of how state-linked and independent actors are trying to weaponize a frontier model across the full spectrum of harm categories the company screens for, from weapons research to live conflict-zone software. Anthropic frames the disclosures as evidence its detection systems are catching this activity before it causes damage, but the findings — especially the acknowledgment that current Claude models can offer meaningful bioweapons-relevant help — add concrete detail to the argument that frontier-model safeguards need to keep pace with capability, a debate that has intensified across the industry as agentic and reasoning capabilities have advanced through 2026.