Anthropic said it identified and disrupted malicious attempts to use its Claude AI across seven threat categories between December 2025 and August 2026. Its September 2026 threat intelligence report covers cases from state-sponsored cyber espionage to research linked to biological and conventional weapons development, the BBC reported.
The cyber cases are the most operationally specific. A group consistent with Russia-linked Midnight Blizzard built AI workflows that detected when their malware was flagged and rewrote the code until it evaded detection. They also compromised the hotel Wi-Fi networks of at least three hospitality vendors through DNS hijacking and targeted government officials and defence industry workers.
ShinyHunters affiliates moved from initial access to bulk data theft in as little as two to three hours at one target. They exfiltrated more than a terabyte of data from a technology provider. At an airline, they accessed tens of millions of passenger records.
A separate case showed a further dimension. One hacktivist on stolen API keys breached at least 14 of 42 targeted organisations. They built a mass-doxxing platform covering tens of millions of records.
Across the cases, Anthropic drew a consistent conclusion. "The main distinguishing feature between these classes of actors is no longer sophistication but intent," the company said. AI has given individual operators capabilities once limited to state-sponsored groups.
That finding arrives at an awkward moment for the cyber market. Moody's flagged cyber as one of the most pressing corporate exposures on its Insurance Emerging Risk Radar this week, warning AI is compressing attack timelines. Lockton's data found average premiums fell around 11% in 2025 even as incident frequency climbed.
The Anthropic report provides primary-source evidence that attacker costs are falling while the threat environment is not improving. That divergence now rests on documented case studies rather than projections alone. The combination is one underwriters advising on cyber placements will need to account for.
Anthropic also documented five cases of users attempting biological research that could support weapons development. Cases included a grant application for gain-of-function research on chikungunya virus transmissibility and work on avian influenza mammalian adaptation. Anthropic blocked all five and withheld the names of institutions involved.
The company acknowledged it could not always determine whether the research was legitimate or malicious. The same queries that could support weapons can also support vaccines and treatments. That dual-use ambiguity is not new to life sciences underwriters already adapting to AI-driven diagnostics and novel drug development.
AI changes its practical reach. Frontier models can now assist with advanced biological research at scale. Questions about what a life sciences liability policy covers when AI is the research tool are ones the sector will need to address.
Finally, Anthropic documented six cases of Claude being used for software linked to firearms, missiles, armed drones and bombs. Cases were tied to users in China, Russia, and Yemen. The report did not identify the Yemeni group by name, but the context points to an Iran-backed non-state actor.
Political violence underwriters have had limited primary-source evidence on AI-assisted weapons development by non-state groups. The Anthropic report provides some. The practical question is whether AI shortens the weapons development cycle for non-state actors. Whether current pricing assumptions reflect that possibility is one the Anthropic cases now make harder to set aside.
