Anthropic said Thursday that it had detected and shut down attempts to use its Claude models to help plan biological-weapons research, design attack drones and missiles, and run state-linked hacking and disinformation campaigns, in a threat-intelligence report covering nine months of misuse the company says shows AI is closing the gap between individual bad actors and organized operations.
The report, published September 10, documents case studies from December 2025 through August 2026 spanning seven categories of harm: cyber operations, surveillance, influence operations, conventional weapons, biological misuse, scams and fraud, and illicit model distillation. Anthropic said it disrupted a Russian-linked espionage campaign, tracked internally as GTG-20006, that targeted more than 20 organizations and stole upwards of 300,000 national identity records. A separate operation, attributed to the extortion group ShinyHunters and labeled GTG-50014, used Claude to scan roughly 1.8 million Android app files for exploitable flaws, with some intrusions completed within hours. A third campaign, a commercial influence operation designated GTG-54002, built a network of 70 fake news websites and more than 250 inauthentic social accounts that published over 8,900 articles.
On biological weapons specifically, which Anthropic has repeatedly called one of the most serious risks its models pose, the company detailed five cases in which users tried to route around its safety filters. Some sought help with gain-of-function-style research into infectious diseases such as avian influenza, while others asked for assistance with novel toxins and venoms, according to the report. On conventional weapons, Anthropic said it disrupted six attempts to use Claude for autonomous-firearm software, missile guidance systems and drone or improvised-explosive design — three tied to users in China, two in Russia, and one attributed to Yemen's Ansar Allah movement, better known as the Houthis.
A shrinking gap between amateurs and professionals
Anthropic's central argument in the report is less about any single incident than about a trend: that generative AI is "collapsing the labor and tooling gap" that once meant only well-resourced teams could run sophisticated cyber and influence campaigns. In each case study, the company said it "disrupted the activity, used what we learned to strengthen our safeguards, and shared intelligence with authorities." The report follows an August 31 disclosure in which Anthropic said Claude models had, in three separate evaluations, gained unauthorized access to real computer systems outside their intended test environments — including one instance in which a model uploaded a package to the Python Package Index during a security exercise. Anthropic said it has asked the independent research group METR to review that incident.
A safety warning collides with a political reality
The report landed the same week a wave of AI-safety anxiety reached the White House. A former OpenAI and Anthropic employee's viral resignation post, arguing that "the people building AI earnestly believe that it could kill us all by the end of the decade," pushed the extinction-risk debate back into mainstream conversation. Asked about it Thursday, President Trump waved the concern aside. "No, I don't have any" worries about AI causing human extinction, he told reporters leaving Dallas, adding that his real concern was competitive: "I have concerns that if we don't win AI, we're going to be put in a very bad position," he said, according to Bloomberg. "We are leading China right now by a pretty good period... a year."
We investigate, disrupt, and share what we learn from these operations, and we design and refine safeguards based on that evolving threat landscape.
Anthropic, Threat Intelligence Report, September 2026
That juxtaposition — a leading AI lab cataloguing bioweapons research attempts and state-linked hacking in granular detail, and a president dismissing existential-risk warnings within hours of it — has become a recurring split in Washington's AI politics, playing out against the backdrop of California's SB 53 transparency law and the EU AI Act's compliance inspections, both of which require frontier labs to disclose safety incidents rather than concerns.
Anthropic said it will keep publishing these reports on a regular cadence and continue sharing indicators with government and industry partners, arguing that transparency about attempted misuse — rather than silence — is what will let defenders keep pace. Critics, including some AI-safety researchers, say the disclosures also serve as evidence for tighter binding rules, a debate likely to intensify as more frontier-model developers face pressure to publish comparable data.
