Tuesday, Nov 17 | --:--
Back to home

Google Finds Attackers' First Fully Automated Exploit Pipeline, Warns of Autonomy Soon

Google Threat Intelligence Group said on November 16 that, for the first time, it identified threat actors using LLM frameworks to fully automate vulnerability discovery and exploitation without a human in the loop, via a multi-agent system it calls exploit_agent found on an exposed staging server aimed at French media, entertainment and education sites. GTIG also reports wider use of open-source agentic pentesting frameworks such as ARTEX and CAIRN, and a resale market for stolen LLM access. BleepingComputer reports Google expects sophisticated, fully autonomous attacks within six months, a projection rather than an observed event.

5 min read Mountain View View as Markdown
A player piano stands in an empty room with its paper roll running dark and no one at the keys, with an official Google logo paper clipping tucked under a torn sage paper flap on the wall beside the piano.

The line Google says has been crossed is narrow but specific: a criminal pipeline that read software patches, worked out what they fixed, wrote an exploit and fired it, with no person in between. Google Threat Intelligence Group (GTIG) says that in the past quarter it identified threat actors using LLM frameworks "to fully automate vulnerability discovery and exploitation without a human in the loop for the first time." The bigger headline, that fully autonomous attacks are months away, is Google's forecast, and should be read as one.

What GTIG found#

The evidence is an exposed staging server hosting a multi-agent framework GTIG calls exploit_agent, which integrates several frontier and local models. According to the report, it continuously compares popular web application code before and after patch announcements, identifies likely vulnerabilities, drafts exploits and executes them automatically. The server also held exfiltrated victim data and scripts for reconnaissance, brute-forcing and installing webshells. Victim telemetry suggested the operator targeted French websites in media, entertainment and education. Russian-language comments in the code refer to it as an "exploit_agent pilot".

GTIG puts this in a wider shift. Attackers are running open-source multi-agent pentesting frameworks in live intrusions, including ARTEX, CAIRN, CyberStrikeAI and PentAGI, and GTIG says it saw CAIRN used against targets including the US aerospace sector. That fits the FBI's warning last week that ARTEX is automating intrusions. The report also says zero-day exploits averaged 11 a month in 2026 against eight in 2025, before rising to 22 in August.

Two other findings matter for anyone running AI systems. GTIG describes "LLM Mixers", underground storefronts reselling access to models through pools of stolen credentials, at prices it observed from $0.10 to $25 per million tokens. And it says an espionage actor targeting cryptocurrency firms altered a coding assistant's configuration so a backdoor downloaded at launch; BleepingComputer identifies the tool as Claude Code and the backdoor as SOMBERMEME.

The forecast#

BleepingComputer reports that Google expects unsophisticated attacks needing "little to no human interaction" within the next three months, and "sophisticated, fully autonomous" agentic attacks within six. We could not find that timeline in the report's text; BleepingComputer attributes it to Google and shows it in a chart from the report.

Limits#

  • This is one vendor's telemetry and interpretation. GTIG does not name the operator behind exploit_agent, quantify its victims, or say whether its exploits succeeded at scale.
  • "First time" means first time GTIG identified it, not proof that no earlier case exists.
  • The three- and six-month timeline is a projection reported by BleepingComputer, not an observed event.

Sources#

Prior Coverage

Earlier Times of AI reporting on this thread.

Scroll to continue reading