Anthropic found Russian and Chinese threat actors using its artificial intelligence chatbot Claude as a weapons engineering assistant, helping develop drone software, anti-torpedo weapons systems, and more.
The extensive report, published Thursday, also detailed activities by a Yemen-based weapons engineering cell and an unspecified Iran-nexus threat actor. The latter used Claude to collect targeting data on US naval forces.
The findings provide examples of a frontier American AI model being used to advance adversarial military capabilities.
Anthropic’s report on the misuses of Claude, from cybersecurity and bioweapons research to engineering autonomous military drone swarms, comes as the Trump administration and Silicon Valley navigate the development of AI amid potential security and safety concerns. Anthropic said it identified problematic activities through its own investigations, banned the accounts involved, and strengthened safeguards.
According to the report, a small, specialized group of Russian actors tied to a regional university, not a Russian state entity, “used Claude to build the core software system, including the drones’ shared swarm memory and fault-tolerant coordination logic,” as well as code for attack, observation, and return to base behaviors, a terminal guidance system for steering drones to their targets, and additional elements for using drones against enemy operators.
Anthropic said that the group built software that could select targets and order attacks without human input, and then they tested it on actual hardware rather than solely in simulations.
“The actors designed the platform for autonomous lethal engagement,” Anthropic said, adding that “the onboard model could select targets” as well as “issue detonation commands without a human in the loop.”
The Russian actors, identified by Anthropic as a specialized “freelance” operation, also developed a tool and trained it on scraped Ukrainian combat footage to determine enemy and friendly assets, and used a fixed coordinate in the Donetsk area of Ukraine as a demonstration strike point.
Anthropic said it identified and investigated the nine accounts related to this work. The “actors circumvented our geographic access controls by routing traffic through commercial virtual private servers.” The threat actors, the company said, created their accounts between late 2025 and early 2026 before starting operations in May.
The Anthropic report also called attention to a China-based threat actor that used Claude in work on an anti-torpedo weapons system, aiming to produce a proposal for the Chinese navy.
They used Claude to write Chinese-language technical specifications for an anti-torpedo fire control system, put together a more than 200-page proposal, compile an executive briefing deck, compare the design against specific US programs using publicly available information, and refine drafts by having Claude role-play as a hostile reviewer.
“They then generated a Chinese-language briefing on US Navy systems from open-source reporting,” the report said. Anthropic discovered the activity as part of internal investigations into suspected weapons development and banned the account.
In another case, Anthropic found a China-based user relying on Claude to gather intelligence on directed-energy weapons.
Separately, Anthropic said it uncovered a weapons engineering operation in northern Yemen, an area where the Iran-backed Houthis control large swaths of territory, that used Claude Code to do the work of human software engineers. Their sessions were focused on guidance, navigation, and control software for rockets and missiles. The company said that these threat actors used Claude to run simulations and diagnose failures.
“We do not have evidence the actors succeeded in fielding an operational device; but they did test-fire a guided rocket,” Anthropic said. “This field test appears to have failed: within hours, the actors returned to Claude to work out why it failed.”
In another instance, Anthropic identified an unidentified Iran-nexus threat actor using Claude to collect and analyze publicly available information to come up with targeting recommendations against US forces.
“The compiled material included a roster of US personnel scraped from captions on public military photographs,” the company revealed, as well as “publicly accessible ship and aircraft transponder identifiers; commercial satellite-imagery query scripts; and an inventory of public websites that exposed US naval movements.”
The company said that it banned the account, shared the information with government authorities, and developed detections for future misuse.
Anthropic’s report comes days after researcher Jacob Coxon resigned from the company in protest of the AI industry’s race to continue developing models, warning in an X post that “the people building AI earnestly believe that it could kill us all by the end of the decade.”
Read the full article here
