Russian hackers taught AI to automatically dodge antivirus software

Anthropic, the company behind the Claude AI model, has uncovered a large-scale cyberespionage operation in which attackers used artificial intelligence to automatically rewrite malicious code every time it was flagged by antivirus tools. According to the company, the attack affected more than 20 organizations — government agencies, defense and intelligence bodies, embassies, think tanks and defense-industry enterprises — mostly in Ukraine and Europe, with additional targets in the Middle East and Asia.

How the system worked

The scheme described by Anthropic was cleverly engineered: an AI system continuously monitored whether antivirus products were detecting the attackers' malware. As soon as one of the tools got "caught," AI agents automatically rewrote and rebuilt the code — repeating the process until the updated version of the malware slipped through undetected once again.

The most alarming part of the story is that this process ran with almost no human involvement. The moment an antivirus program "recognized" the malicious software, the AI would almost instantly find a way to bypass protection again. Once successfully reworked, the tool was prepared for real-world deployment via temporary servers, where it was used for phishing, ClickFix-style attacks and DNS record manipulation.

Not one virus, but an entire arsenal

This wasn't a single piece of malware but a whole toolkit: two types of Windows implants, a tool for breaking into mobile devices, a program for stealing passwords stored in browsers, a phishing platform disguised as a government organization, and an administrative console for managing compromised accounts. According to Anthropic, all of these tools were continuously refined with the help of AI, and automated workflows covered a significant portion of the operation — from malware development and infrastructure takeover to phishing, maintaining access to compromised systems, and exfiltrating data.

Humans didn't disappear from the process entirely, though: according to the investigation, specialists fine-tuned the capabilities used within Claude Code whenever precise adjustments to the whole system were needed. But routine operations after that point could run automatically.

Why this is dangerous

This is precisely where Anthropic sees the main threat. Cyberattacks typically follow a cycle: an attacker builds a tool, defensive systems detect and block it, and the attacker then has to develop a new version — a process that takes time, resources and skilled specialists. AI can drastically shorten this cycle: if a system itself tracks exactly what defenses have flagged and immediately rewrites and retests the tool, the whole process speeds up dramatically. The company warns that classic static detection methods may no longer be able to effectively raise the cost of attacks for adversaries the way they once did.

According to the report, Ukraine, military drones and supply chains were recurring themes in the attackers' correspondence. Anthropic links the entire operation to Russian espionage — specifically to the state-sponsored group known as Midnight Blizzard. The investigation found that one of the operators used the Russian language and operated under the alias JackPoterz.

This case is just one of many episodes described in Anthropic's new comprehensive report on AI misuse. Besides cyberattacks, the document covers cases of malicious influence operations, surveillance of individuals, fraud, biological research, conventional weapons development, and unauthorized attempts to gain access to AI capabilities. The company stresses that these are not random examples of everyday chatbot misuse, but cases hand-picked by its specialists that represent a significant or fundamentally new security risk.

Read also: Mobile operators in Czechia: which one to choose in 2026

Source: novinky.cz