Wired: новый вредонос-червь атакует ИИ-инфраструктуру, крадёт логины и уничтожает файлы жертв

In July 2026, Wired described a new type of malware that worms its way into AI coding systems: the program steals developers' data and credentials, and on the attacker's command activates a "death switch" — destroying files and locking legitimate users out.

How the new malware works

The malware behaves like a network worm but targets AI development infrastructure. According to Wired, it penetrates deep into AI coding systems and entrenches itself in victims' "blind spots" — parts of the infrastructure that security teams barely monitor. From there, the program quietly harvests data and accounts, expanding its access to developers' environments.

AI coding systems include assistants and autonomous agents that write, edit and run code on a developer's behalf. Such tools operate with broad user privileges, so malware entrenched in them inherits the same capabilities — reading, modifying and deleting everything the agent can reach.

Credential theft in an AI development environment is especially dangerous. Login data opens the way to code repositories, model API keys and cloud services: by compromising a single tool, the attacker gains not an individual computer but a company's entire development perimeter.

What the "death switch" is

The "death switch" is a destructive function that the malware's operator can activate at any moment.

"The malware is capable of destroying files and locking real users out of the system," the

Wired report says.

This scenario turns espionage into sabotage: the victim simultaneously loses both their data and control over their own infrastructure. Mechanically, it is closer to wipers — data-erasing programs used in attacks on critical infrastructure — than to classic ransomware, which encrypts files for a ransom.

Why AI infrastructure became a target

AI programming assistants are granted the broadest privileges inside companies: access to source code, secrets, CI/CD pipelines and cloud environments. The more autonomy AI agents have, the more attractive they are as a target — compromising one tool opens up everything it "sees". Monitoring, meanwhile, lags behind: traditional security tools struggle to understand what AI agent behavior should count as normal, and it is precisely these blind spots, according to Wired, that the new malware exploits.

Attacks on the AI development chain became a notable trend in 2025–2026: security researchers regularly find malicious packages and extensions aimed specifically at developers and their tools. The new worm is the next step in the same logic: attack not the model, but the infrastructure around it.

What this means

AI infrastructure security is turning into a discipline of its own: AI coding tools are now so widespread that specialized malware is being written for them. Companies deploying AI agents should apply the same rules to them as to human developers: least privilege, key rotation and monitoring for anomalous activity. There is a separate takeaway for teams that give agents access to production: a destructive function like the "death switch" means that backups and environment isolation are no longer a formality.