AI Was Supposed to Choose Its Own Targets to Kill: Russia Was Developing a Swarm of Kamikaze Drones

Anthropic, the company behind the Claude artificial intelligence model, has discovered that a group of Russian freelancers used its AI to build an autonomous swarm of kamikaze drones capable of independently selecting a target and issuing a detonation command without any human involvement. This is according to a company report on AI misuse.

Anthropic assigned the project the internal code name GTG-27005, while its participants referred to the operation as "DronDoc" or "Seraphim." According to Anthropic, this was not a state-run operation but rather a small team, reportedly Russian freelancers. Nevertheless, the participants themselves claimed their work was funded by Russia's Advanced Research Foundation, the National Technology Initiative, and the Ministry of Defense.

How AI Was Involved in the Development

According to the report, the developers used the Claude Code tool to write and test software, and also relied on simulations and rented computing power to train their models. AI was used to build the system's core software — designed to enable multiple drones to communicate, coordinate with one another, and carry out other autonomous functions. The model also helped write code for guiding drones toward their targets.

The critical turning point, in Anthropic's assessment, came when the project moved beyond simply assisting programmers and toward autonomous decision-making about attacks. The drone's onboard model was designed to independently select a target — with categories that reportedly included "human" — and to issue a detonation command with no human in the decision-making loop.

Training the System to Recognize the Enemy

Investigators found that the group's work wasn't confined to computers: they uploaded low-level software onto real development boards, prepared single-board computers, and physically connected different components of the system. This doesn't mean such an autonomous system has already been deployed on the battlefield, but it's clear the developers were working toward exactly that outcome.

Part of the effort also involved computer analysis of footage from the war in Ukraine: the group trained a computer-vision system on real combat footage so it could distinguish between "enemy" and "friendly" categories, with Russian military equipment placed on the list of approved objects. Investigators also noted that the individuals involved repeatedly used specific coordinates in Donetsk region as a demonstration attack point, while the geography of planned missions included frontline towns and corridors within Ukraine.

Not a State Project, But State-Funded

Although the project concerned military technology and its creators, according to Anthropic, operated from Russian territory, the company does not classify the group as a state entity. Instead, it describes them as "a small, specialized team of freelancers" who juggled both civilian and military contracts. In total, nine accounts linked to the group were identified, eight of which were used solely for ordinary freelance work unrelated to weapons software development. The investigation also pointed to possible connections with a regional university and a federal research center affiliated with the Russian Academy of Sciences.

Anthropic notes that Claude was not the team's only tool: alongside it, the group used its own project files, simulation environments, and computing power to train models. In other words, Anthropic's AI model was just one element within a much broader technical infrastructure behind the project.

Source: novinky.cz