Anthropic says a Russia-linked group used Claude between December 2025 and August 2026 to assist a cyber-espionage campaign targeting Ukrainian government, military and diplomatic organizations, European bodies and companies connected to the drone supply chain. The company says the activity involved more than 20 organizations and that it disrupted the operation and shared intelligence with authorities and industry partners.
The activity was identified by Anthropic as GTG-20006. Claude reportedly helped with technical work across the campaign, but human operators supplied the objectives, infrastructure and intent.
Who the campaign reportedly targeted
The reported targets included Ukrainian state organizations, military and diplomatic personnel, European organizations, defense bodies and companies involved in the drone supply chain. More than two dozen Ukrainian state organizations were reportedly scanned across email and remote-access systems.
At least two drone-component manufacturers reportedly had their mailboxes bulk-exported. At least eight organizations, including a national prosecutor’s office and a military education institute, were reported to have had cloud email stolen.
| Target area | Reported activity | Reported scope or material |
| Ukrainian government, military and diplomatic organizations | Scanning of email and remote-access systems | More than two dozen Ukrainian state organizations |
| Drone supply chain | Mailbox theft and technical-data collection | At least two manufacturers; a proprietary drone-vision software development kit was reportedly stolen and analyzed |
| Hotel guest Wi-Fi | Redirection of travelers toward attacker-controlled infrastructure | At least three guest-Wi-Fi vendors were reportedly compromised |
| Broader campaign | Planning and live operations | More than 20 organizations; hundreds of gigabytes of data were reportedly stolen |
What Claude reportedly did
Anthropic’s account describes Claude as an accelerator for specialist cyber work. The reported uses included reconnaissance, phishing infrastructure, command execution, credential harvesting, organization of stolen data and modification of malware.
Anthropic said automated systems checked whether malware implants had been detected and modified flagged artifacts for redeployment. That workflow could shorten the time between a security product identifying an implant and an operator preparing another version.
The reported campaign therefore involved AI assistance inside a human-directed operation. Claude did not set the targets or create the campaign’s objectives independently; operators provided the surrounding infrastructure and direction.
The drone-software theft and hotel Wi-Fi route
The drone-related operation reportedly included the theft of a complete proprietary software development kit for a drone-vision system, followed by analysis of that material. The reported account does not attach a quantified battlefield or commercial outcome to the theft.
A separate part of the activity involved hotel guest Wi-Fi. Operators reportedly compromised at least three vendors and redirected travelers toward infrastructure controlled by the attackers. That route could expose visitors to credential theft or malware delivery when they connected through affected networks.
What Anthropic says happened to its own systems
Anthropic said its systems were not breached during a separate operation that sought an unreleased Claude model. In that operation, production API keys were reportedly taken from a compromised testing system at another AI company.
Anthropic said it disrupted the malicious activity, strengthened its safeguards and shared intelligence with authorities and industry partners where appropriate. The company’s disclosure places the security challenge beyond model refusals alone: defenders also have to monitor accounts, infrastructure and the way automated tools are incorporated into live intrusion work.
