Anthropic has revealed a cyberespionage campaign linked to the Russian group known as Midnight Blizzard, which has been exploiting AI technology to enhance its malware evasion tactics. The company’s latest threat intelligence report, published this week, details the activities identified and halted between December 2025 and August 2026.
AI Utilized for Malware Evasion
According to Anthropic, Midnight Blizzard employed Claude AI to assess the effectiveness of its malware against security tools. When detection occurred, AI agents autonomously adapted and redeployed the malware, continuing this process until it bypassed security measures once more. This innovation shifts the burden of the detection-evasion cycle onto defenders, allowing attackers to refine their tools more swiftly than traditional methods allowed.
The report highlights that over 20 organizations were targeted, including Ukrainian and European governmental entities, defense and intelligence agencies, and think tanks. The group’s reach also extended into the Middle East and Asia, indicating a broad spectrum of espionage targets.
Specific Targets and Methods
Midnight Blizzard’s operations included the theft of sensitive data, such as mailboxes from drone component manufacturers and a proprietary software development kit for a drone vision system. Detailed examination of this data involved reverse engineering and analysis of the hardware and supplier dependencies.
The group also infiltrated hospitality services by compromising hotel guest Wi-Fi systems, using DNS hijacking techniques. This method was previously documented by Microsoft under the alias CaptiveCrunch. Additionally, they manipulated WhatsApp accounts of high-profile Ukrainian figures, utilizing headless browsers to suppress read receipts and export conversation data without detection.
AI Infrastructure as a Target
Anthropic’s report also identifies a trend where threat actors are targeting AI infrastructure and credentials. Notably, a group labeled GTG-50021 established a fraudulent Claude AI reseller service, capturing users’ credentials through a proxy service.
Another group, GTG-50020, focused on financial gain by extracting API keys from AI vendors, subsequently targeting approximately 30 AI companies. Their objective was accessing a pre-release Claude model, though these attempts were unsuccessful.
Anthropic emphasizes the need for organizations to safeguard AI credentials as rigorously as they protect production credentials, given their potential misuse in cyber operations.
These findings form part of a broader examination of AI misuse, covering areas such as influence operations and weapons development, underscoring the growing complexity of cyber threats in the digital age.
