Google catches PROMPTFLUX, an AI malware that rewrites itself

Traditional antivirus software works a lot like airport security checking passports. If your digital fingerprint is on the blocklist, you get stopped at the gate.

Now imagine a piece of malware that prints itself a completely new identity every few seconds.

Google's Threat Intelligence Group (GTIG) just caught a new threat doing exactly that. It is called PROMPTFLUX, and it might be one of the clearest signs yet that malicious software is getting an AI upgrade.

A shapeshifter in real time

PROMPTFLUX uses the same generative AI technology behind models like Google Gemini to rewrite its own code on the fly.

Standard malware runs on fixed, predetermined scripts. Even sophisticated variants usually follow a rigid playbook once inside a network.

PROMPTFLUX takes a very different route. It learns and adapts in the middle of an attack, generating new techniques as it moves through a system.

Its sole objective is to constantly alter its own code so endpoint defenses and signature-based scanners have zero chance of matching a known pattern.

Caught in the lab

The good news is that GTIG caught it early. The samples analyzed contained unfinished components, suggesting the software was still in testing rather than active deployment.

Google confirmed that PROMPTFLUX has not yet been seen infecting targets in the wild, and the company has already neutralized the accounts and assets tied to the project.

However, Google warned that the looming danger is commercialization. If tools like this hit illicit underground markets, even low-skilled attackers could soon deploy advanced, shapeshifting malware.

To prepare, Google is rolling out defensive frameworks built specifically for AI architectures, alongside counter-AI initiatives like its Big Sleep system, designed to hunt and patch zero-days before automated attackers exploit them.

Why it matters

We think PROMPTFLUX marks the official end of traditional signature detection as a viable primary defense line.

For everyday users and security teams, static blocklists are officially on life support. If malware can mutate programmatically at runtime, defenses have to evaluate live intent and behavior in real time, which inevitably makes security stacks heavier and more complex.

For the cybersecurity industry, the arms race is no longer human versus software. It is AI versus AI.

Using an LLM to generate continuous polymorphic mutations is clever because it exploits the structural latency of human patch cycles. But lowering the barrier to entry through dark-web distribution could create widespread instability across enterprise infrastructure. Digital safety is rapidly coming down to whose model can out-reason the other.