technology 6 min read

Google Says Russia Ran First Fully Automated AI Cyberattack

Google has documented the first end-to-end AI-driven cyberattack, with no humans steering any stage. It is not just a sharper tool—it is a new kind of warfare, and the rules do not exist yet.

  • Artificial Intelligence
  • Autonomous Weapons
  • Cybersecurity
  • State-Sponsored Hacking
  • Google Cloud

The moment the loop closed

Google announced last Monday that its Threat Intelligence Group and incident-response arm Mandiant had identified the first cyberattack run completely end-to-end by artificial intelligence. There was no operator waiting between vulnerability detection and code injection, no human confirming a target or adjusting a payload once the system was running.

That may sound like a small engineering milestone. It is not. It is the point where autonomous weapons entered the cyber domain in the way military analysts have warned about for years.

Attribution rested on Russian-language strings in the operation’s source code. A configuration file instructed users to paste API keys into a specific field rather than passing them through a chat interface—an ergonomic fingerprint that suggests someone built this deliberately, not as a sloppy script.

Victim data found on a staging server pointed primarily at French websites in media, entertainment and education. Google did not name a state or affiliate, which is significant. The decision to withhold attribution signals either ongoing investigation or a deliberate choice not to escalate publicly—a pattern familiar from Russian cyber operations going back at least a decade.

What matters more than the missing name is the architecture underneath it.

How exploit_agent works

Google’s researchers named the centerpiece “exploit_agent.” It chains multiple AI agents together with frontier language models and locally hosted systems. The pipeline does three things continuously: it reads software updates published by vendors, compares old and new code to infer what changed, then drafts and executes an exploit automatically.

That inference step is the breakthrough. Earlier AI-assisted hacking relied on humans feeding prompts, selecting targets, or reviewing output. Here the system observes a patch, deduces the vulnerability, writes the exploit and fires it—all in sequence, unattended.

The same operation included scripts for internal reconnaissance, password guessing and webshell deployment. One capability stood out separately: the ability to detect and erase rival attackers’ webshells. That is not defensive. It is competitive offense, a tool for monopolizing access after a breach.

Google said it has now identified threat actors who use such frameworks to fully automate vulnerability discovery and exploitation without a human in the loop for the first time.

The numbers behind the shift

Speed is the most visible advantage, and the data supports it.

After two PHP vulnerabilities were patched on October 21, hundreds of purported exploits appeared within days. Most were nonfunctional, likely generated by AI with minimal filtering. But the signal is unmistakable: AI-drivated exploration turns a patch release into a flood of candidates almost instantly.

Zero-day counts tell an even starker story. Google recorded twenty-two exploited zero-day flaws in August alone. Earlier in 2026 the average sat at eleven per month. In 2025 it was eight. The trajectory is steep and accelerating.

Exploit generation was never the bottleneck in traditional hacking. Skilled operators take weeks to reverse-engineer a new vulnerability. An agent system reduces that window to hours or minutes. Defensive timelines—that gap between disclosure and widespread exploitation—compress accordingly.

It is not only Russia

Google flagged several other developments in the same report. Open-source offensive frameworks called CAIRN and PentAGI surfaced in activity targeting a U.S. aerospace organization and entities across Asia. A China-nexus group appeared to be constructing its own automated attack framework, likely designed for persistent espionage rather than destructive operations.

These are separate from the Russian-sourced exploit_agent operation, but they share the same logic: automate the tedious parts of intrusion so humans can focus on strategy. The effect multiplies. More actors can scale lower-tier intrusions to levels previously reserved for well-resourced teams.

Commercial companies are moving in the same direction. Anthropic released Mythos in April, a model explicitly tuned for cybersecurity tasks. An OpenAI agent later demonstrated autonomous hacking capability against the development platform Hugging Face during a summer incident, though that activity did not target specific external organizations.

The ecosystem is shifting faster than the policy response.

Why attribution becomes harder, not easier

The conventional thinking assumes autonomous tools make attribution simpler because machines leave clean digital footprints. That is incomplete.

Automated exploitation generates far more noise. Floods of failed attempts, randomized payloads, and overlapping indicators from competing tools make it harder—not easier—to distinguish one operator’s signature from another’s. When multiple groups deploy variations of the same open-source framework, as Google has observed with CAIRN and PentAGI, attribution degrades across the board.

Moreover, the Russian-language comments in exploit_agent suggest careful engineering. Operator discipline inside the toolchain matters less when the system runs autonomously, but the people who build and maintain those systems still leave traces. The challenge is that those traces now arrive in greater volume and from more sources simultaneously.

This is precisely why Google stopped short of naming an affiliation here. The pattern of withholding attribution despite confident technical indicators suggests the company judges the evidentiary standard higher—or perhaps recognizes the political cost of accusing a nuclear power without ironclad proof.

Who wins, who loses

The immediate winners are anyone who can afford to run autonomous offensive tooling at scale. Small and mid-sized criminal groups gain capabilities that previously required dedicated talent. State actors gain deniability and speed. Both benefit from a shrinking window between patch publication and active exploitation.

The losers are organizations relying on manual security monitoring and human-driven incident response. A defensive team that reacts to alerts cannot compete with an agent that finds, exploits and installs persistence in minutes. Even mature security operations centers that assume their detection coverage is adequate will struggle when attack velocity consistently outpaces investigation capacity.

France is already feeling the pressure, given the targeting profile Google described. But the vulnerability landscape is global. Any organization running widely deployed software with remotely exploitable bugs is exposed, regardless of geography.

What the absence of rules reveals

The nuclear arms control infrastructure emerged because both sides understood the consequences of unregulated escalation. Cyber does not yet have that clarity.

When an attack is fully automated, the chain of command between decision and execution vanishes. Deterrence theory depends on credible attribution and proportional response. Autonomous systems erode both: they generate ambiguity faster and multiply the number of potential perpetrators involved in any single campaign.

The United States and Russia have discussed AI governance bilaterally, but no framework addresses automated offensive cyber capabilities specifically. Existing norms around responsible AI focus on generative content, alignment safety and economic disruption. They say nothing about machine-speed intrusion pipelines or the legal status of autonomous exploit agents acting without human authorization.

That gap is the real story behind Google’s report. The technology is here. The norms are not.