Google says a suspected Russian-speaking attacker used an AI agent to turn software security patches into working exploits with no human involved, which it calls the first known case of its kind. The Google Threat Intelligence Group (GTIG) documented the activity in its fourth-quarter 2026 AI threat tracker, published on 16 November 2026. Days earlier, the FBI warned that an open-source AI hacking tool called ARTEX has been used to break into networks in under three hours.
Key takeaways
- The framework, called "exploit_agent," compared code before and after security updates, found the flaw the update fixed, wrote an exploit and ran it automatically.
- Google found it on an exposed staging server and says the attacker then stole data from victim organisations.
- The FBI says ARTEX can take an intruder from first scan to privileged access in under three hours.
- The practical lesson is that a published patch no longer buys defenders much time.
How the exploit_agent AI hacking framework worked
According to The Register and other coverage of Google's report, exploit_agent is a multi-agent system for vulnerability research and weaponisation that connects to several frontier and local AI models. It used Russian-language tooling and was set up for continuous "differential analysis" of popular web application software, meaning it examined what changed in each update to find the vulnerability being fixed. It then built and launched exploits against sites that had not yet applied the update.
Memeburn reports Google's evidence of origin is mainly Russian-language code comments, including a setup file telling users to enter AI account keys for "auto-exploitation." Google has not named a group or country. Victim data pointed to French media, entertainment and education websites, which suggests under-maintained sites rather than banks or governments.
FBI warning about the ARTEX tool
The FBI's advisory of 13 November describes ARTEX as an autonomous penetration-testing framework written in Go. It uses AI agents to scan for weaknesses, exploit them, escalate privileges and steal data with little human input, and is especially effective on overlooked systems such as partner portals, legacy services and third-party integrations. Researchers had found more than 1,600 publicly exposed ARTEX consoles by early November, according to BleepingComputer, and the FBI cited a supply-chain breach at ConnectWise that reached Microsoft 365 customer environments. Google lists other AI offensive frameworks in use: CAIRN, CyberStrikeAI, HexStrike AI and PentAGI.
What defenders should do now
- Shorten patch windows. Apply security updates as soon as they are released, especially to internet-facing web software.
- Watch for signs of compromise. Monitor for webshells, brute-force attempts and unusual admin logins.
- Segment networks so one compromised system cannot reach everything.
- Check forgotten systems: partner portals, old services and third-party integrations.
- Keep incident response plans current.
Google says these frameworks are noisy and leave traces, which makes them easier to detect, but expects them to need less human help and become stealthier.
Frequently asked questions
Was this a fully automated AI cyberattack?
Google says it is the first known case of an AI agent turning patches into working exploits with no human in the loop. Google has not publicly named the attacker.
Does this affect my website?
Sites that run popular web software and delay updates are the most exposed. Patch quickly and ask your host about update status.
Is ARTEX illegal?
It is an open-source penetration-testing tool. The FBI warns that criminals have used it in real attacks.
For related stories, see our AI cybersecurity hub and our report on OpenAI's AI-built Linux exploits.
Sources: Google Threat Intelligence Group Q4 2026 AI threat tracker, via The Register, Memeburn, Mezha and Osegi; FBI advisory via BleepingComputer; Secure.com.