Russian developers used Anthropic’s Claude artificial intelligence to develop software designed to coordinate an autonomous swarm of strike drones. This is stated in the company’s report on detected cases of misuse of its models.
According to Anthropic, the project involved a group the company labelled GTG-27005. The developers worked with Claude Code, using it to write, test and debug drone software. The project, which appeared under the names DronDoc and Serafim, began to be implemented around May 2026.
To gain access to services despite geographical restrictions, the project participants used commercial virtual servers.
One of the key tasks was to create a system for interaction between drones. It was intended to enable information exchange between the aircraft and allow the group to continue carrying out a mission even after the loss of one or more drones.
Separately, the developers created a compact language model that was meant to determine the drone’s next action. Possible scenarios included attacking a target, surveillance, or returning to base. Anthropic notes that the system could independently identify targets and transmit the command to strike them without the operator’s direct involvement.
The software also included a target category labelled “person”. To improve computer-vision algorithms, the developers used combat footage from Ukraine. In it, objects were classified as “friendly” or “enemy”.
The project’s documentation and test materials also mentioned Donetsk Oblast as an example of a possible area of deployment. Ukrainian frontline cities and movement routes were considered as potential scenarios for future missions. It is currently unclear whether this actually worked and whether the combat use of such a swarm was confirmed.
At the same time, Anthropic did not establish that the group was directly part of a Russian state structure. In the company’s assessment, it was likely a small team of freelancers carrying out various civilian and military contracts.
Anthropic linked nine accounts to the project. The company also identified participants’ connections to a regional university and a federal research centre associated with the Russian Academy of Sciences.
The project participants claimed they received funding from Russia’s Foundation for Advanced Research, the National Technology Initiative and the Russian Ministry of Defence. Anthropic says it was unable to independently verify these claims. After detecting the activity, the company blocked the accounts linked to the project.
Earlier, Anthropic reported in a report that the cyber actor GTG-20006, which researchers link to the Russian hacking group Midnight Blizzard, used the Claude Code artificial intelligence tool to automate cyberattacks on Ukraine and European countries. The hackers carried out operations targeting military intelligence objectives in the governments of Ukraine and Europe, as well as diplomatic and defence organisations and individuals linked to US foreign policy. The use of AI enabled the perpetrators to bypass traditional cyber-defence systems and autonomously modify malicious software code.