Skip to content
← All stories

Vulnerabilities · AI Security

Google says AI agent built a working exploit from a security patch

Google researchers say an AI agent created a working exploit from a security patch, with no human help. The attack targeted French media, entertainment, and education sites, according to The Register.

Ross Grant2 min read

Laptop installing a security update

The short version

Google researchers say they found an AI system that could turn a security patch into a working cyberattack, all by itself. The AI compared software before and after a patch, found what changed, and used that to break in. The attack targeted French websites in media, entertainment, and education. Only one news outlet has reported this so far, and Google has not published the full details.

What happened?

Google’s Threat Intelligence Group reportedly found that an AI agent created a working exploit from a security patch, without human help. The AI system targeted French media, entertainment, and education websites. The Register is the only outlet to report this so far, citing Google’s Q4 2026 AI threat tracker.

How the attack worked

According to The Register (opens in a new tab), Google’s Threat Intelligence Group (GTIG) found evidence that an AI agent, with no human in the loop, turned a security patch into a working exploit. The AI system did this by comparing code before and after a security update, spotting the change, and then automatically developing and running an exploit. This process, known as differential analysis, is a common way to find out what a patch fixes. In this case, the AI did the whole job itself, from finding the vulnerability to exploiting it.

GTIG found a staging server running a multi-agent framework called exploit_agent. This system integrated several large language models (LLMs) and was set up to scan popular web application software for new patches, analyse them, and weaponise any vulnerabilities it found. The framework was configured with Russian-language tooling and could run post-compromise scripts, such as internal reconnaissance, brute-forcing passwords, installing webshells (malicious programs that give attackers remote access), and removing other attackers’ webshells.

Who is affected

The stolen data on the server suggested the attackers targeted French websites in the media, entertainment, and education sectors. The Register reports that the server contained data from multiple victim organisations, but does not name them or give numbers.

What experts say

Google’s report, as described by The Register, calls this a significant advance in how attackers use AI. It is the first time GTIG has seen a fully autonomous AI agent turn a patch into a working exploit and use it in the wild. The team believes the low barrier to entry and the ability to scale these attacks mean more threat actors are likely to adopt similar frameworks.

The report also mentions other AI-powered offensive frameworks, such as ARTEX, CAIRN, CyberStrikeAI, HexStrike AI, and PentAGI. Both financially motivated criminals and suspected espionage groups have experimented with these tools, including some with links to China (PRC nexus).

What you should do

Organisations should review their patch management processes. Applying patches quickly is still critical, but defenders should be aware that attackers may now use AI to weaponise new vulnerabilities much faster than before. Monitoring for signs of automated exploitation and reviewing logs for unusual activity after patch releases is recommended.

No primary source from Google is available yet. The story is based on The Register’s reporting and should be treated as unconfirmed until Google or another authority publishes the original report.

Claimed vs confirmed

  • Unconfirmed

    An AI agent turned a security patch into a working exploit with no human help.

    Reported by The Register, citing Google’s Threat Intelligence Group, but not yet confirmed by Google or any other primary source.

  • Unconfirmed

    The attack targeted French media, entertainment, and education websites.

    Based on data found on the attacker’s server, according to The Register’s summary of Google’s report.

Why it matters

If confirmed, this is the first known case of an AI agent carrying out every step of a cyberattack, from finding a vulnerability to exploiting it, without human help. This could make attacks faster and harder to defend against, especially right after patches are released.

Key points

  1. Google says an AI agent created a working exploit from a patch, with no human help.
  2. The attack targeted French media, entertainment, and education websites.
  3. The AI used a multi-agent framework to automate every step, from finding to exploiting vulnerabilities.
  4. Google warns that more attackers may adopt these AI-powered frameworks.
  5. The story is based on a single news report and has not been confirmed by Google directly.

Sources

  1. The Register

    Reference

The day's cyber news in 5 minutes.

Cyberend Daily is one short email with the stories that matter, written so anyone can follow.

Includes our free beginner roadmap · No spam · Unsubscribe anytime