Anthropic has disclosed a sophisticated operation in which a Russian state-linked group tracked as GTG-20006 used autonomous AI agents to automatically reconstruct its malware every time security products detected it.
The group maintains documented ties to Midnight Blizzard, also known as APT29 and Cozy Bear. Its workflow formed a fully automated cycle: AI agents continuously evaluated the performance of deployed tools against known defenses and, upon detecting that an artifact had been flagged, initiated independent code changes and recompilation to bypass the specific detection.
Automation extended well beyond the malware itself. The agents registered domains, prepared hosting infrastructure, launched phishing emails, monitored command-and-control channels and tracked implant persistence on victim networks.
Windows implants referenced in the report include PowerChrome, WUEngine, Shadow C2, MiniPlasma and CloudSyncSvc. On mobile platforms the group deployed GiftDrop and GiftsExpress for Android, DarkSword for iOS, a browser password stealer and a dedicated phishing platform.
The campaign ran during July and August 2026 and overlapped with the previously reported CaptiveCrunch operation. More than twenty organizations were targeted, including government ministries, defense agencies, embassies, think tanks and defense-sector companies in Ukraine, Europe, the Middle East and Asia.
In one confirmed breach the attackers accessed over 300,000 national identity records and commercial registration data belonging to more than half a million companies. Anthropic terminated the campaign and released a full technical report, noting that the automation has effectively reversed the traditional cost advantage previously enjoyed by defenders.
