Autonomous AI Cyberattack | How exploit_agent Hit French Websites With No Human in the Loop

Reading Time: 5 min

The first autonomous AI cyberattack may already be behind us. On November 16, 2026, Google’s Threat Intelligence Group (GTIG) reported that Russian-speaking hackers used an AI system called “exploit_agent” to find and exploit software flaws with no human in the loop. The targets were French websites in the media, entertainment and education sectors. Google says it is the first time it has seen an attack chain run end to end by AI agents.

For years, security experts warned that AI would one day hack on its own. That day appears to have arrived quietly, on an exposed server, against mid-sized websites. The bigger news is not who was hit. It is how little skill the operators needed and how fast the tool turned patches into working exploits.

What Google Found

The discovery came from a mistake by the attackers.

An exposed staging server

GTIG researchers found a server that the operators had left open. It held the exploit_agent framework, stolen data from victims and configuration files. The victim data showed a clear focus on French sites.

A chain of AI agents

The framework links several AI agents together. It works with both frontier models and models hosted on the attackers’ own hardware. Each agent handles one stage of the attack, and the chain runs without a person approving each step.

Patch in, exploit out

The system watches popular web software for new security patches. It compares the code before and after each fix, works out what the change protects against and writes an exploit for the old version. Then it runs that exploit against real targets on its own.

Tools for the next stage

Researchers also found tools for deeper access after the first break-in. One tool removes web shells planted by rival hackers. That lets the operators push competitors off a server they have taken over.

The patch you release today is the exploit someone's AI writes tonight.

Who Is Behind the Autonomous AI Cyberattack?

Attribution is limited, and it is worth reading carefully.

Russian-language clues

The configuration files contain comments written in Russian. One tells the user to enter API keys for “auto-exploitation” in the file rather than in a chat. That points to Russian-speaking operators. It does not prove links to the Russian state or to any known group.

What “Russian-speaking” really means

Russian is a working language across much of the cybercrime world, from Russia and Belarus to Ukraine, Kazakhstan and beyond. Language clues help narrow the field, but they rarely settle who ordered an attack. Criminal crews, hacktivists and state units all write code in Russian. Careful analysts treat this kind of evidence as a lead, not a verdict.

A tense backdrop

The report landed the same day the EU sanctioned 30 people and two entities over Russian information manipulation and cyber operations, including the pro-Russian group NoName057(16). France has faced repeated Russian-linked hacking and disinformation campaigns. No one has tied the sanctions to this AI operation, but the timing adds to concern in Paris.

Not the only one

Google also found a China-linked group building a similar autonomous framework, most likely for espionage. The same report describes stolen AI service credentials and tampered settings in AI coding tools. Autonomous offense is becoming a trend, not a one-off.

Why This Changes the Patching Race

Defenders have long relied on a gap between a patch and a working exploit. That gap is closing fast.

Every patch is a roadmap

Security researchers have always diffed patches to find the flaw behind them. It took skill and days of work. An AI agent can now do the same in hours and repeat it for every update across many products at once.

Low skill, high output

Google warns that the barrier to entry is low and attackers will very likely keep using such frameworks. A small crew with API keys can now run an operation that once needed a team of exploit developers. Mid-sized sites that never drew skilled attackers are now in range, because an AI agent does not care whether a target is famous.

Speed leaves traces

There is one piece of good news. Google says these frameworks trade stealth for speed. They leave noisy forensic traces, such as bursts of exploit attempts right after a patch. Teams that watch for that pattern can catch them.

Same tools, both sides

Defenders are using AI too. In the same week, researchers used GPT-6 Cyber to find dozens of Apple flaws before criminals did. The race now favors whoever automates first.

What Website Owners and Security Teams Should Do

Treat the time between a vendor patch and your deployment as live exposure. Every day an update waits in a queue is a day an AI agent can use it against you.

  1. Patch web software within days, not weeks. Content management systems, plugins and web frameworks are prime targets for patch-diffing agents.
  2. Turn on automatic updates where the risk of breakage is low, especially for security-only releases.
  3. Use a web application firewall with virtual patching to cover the gap before you can update.
  4. Watch for post-patch spikes. Alert on surges of unusual requests to recently patched endpoints.
  5. Hunt for web shells. Scan servers for unknown files and scripts, including ones another attacker may have left.
  6. Protect your AI keys. Rotate API keys, restrict them by IP and monitor usage. Our article on exposure-based triage shows how to decide what to fix first.
When attackers automate, your patch window becomes your attack surface.

The autonomous AI cyberattack on French websites was not the largest or most damaging attack of the year. Its importance lies in what it proves. Machines can now find a flaw, build the exploit and use it without a human hacker at the keyboard.

That shifts the advantage to speed. Organizations that patch fast, monitor closely and automate their own defenses will keep pace. Those that patch on a monthly cycle will find AI agents waiting for them. Google’s threat team flagged the same shift inside malware itself — see AI-powered malware: PROMPTFLUX and PROMPTSTEAL.

Want to see how quickly AI-found flaws reach critical software?

Read our deep-dive: Java4Shell CVE-2026-114938.

Close your patch window before an AI agent walks through it.
SECITHUB FAQ banner for the first autonomous AI cyberattack on French websites
What is the first autonomous AI cyberattack?

Google reports that Russian-speaking hackers used an AI framework called exploit_agent to find and exploit flaws in French websites without human involvement.

What is exploit_agent?

A chain of AI agents that compares software before and after security patches, writes exploits for the fixed flaws and runs them against targets automatically.

Who was targeted?

French websites in the media, entertainment and education sectors, according to victim data on the attackers’ server.

Who is behind the attack?

The operators appear to speak Russian, based on comments in the code. Google has not linked them to a known group or to the Russian state.

Can autonomous AI attacks be detected?

Yes. Google says these frameworks trade stealth for speed and leave noisy traces, such as bursts of exploit attempts right after patches.

How can website owners protect themselves?

Patch web software within days, enable automatic security updates, use a WAF with virtual patching, monitor post-patch traffic and hunt for web shells.

Sources

Google says Russian-speaking hackers ran the first fully automated AI cyberattack – CyberScoop

Google says attackers used AI to turn patches into exploits with no human – The Next Web

Google says Russian-speaking hackers ran the first fully automated AI cyberattack, using exploit_agent to turn patches into exploits against French websites. Here is how it worked and what defenders must change.
0 0 votes
Article Rating
Subscribe
Notify of
guest

0 Comments
Oldest
Newest Most Voted