Tuesday, 17 November 2026

AI agents now find and exploit flaws without a human, Google says

Google's threat team says attackers have begun running AI systems that compare patches, write exploits and fire them with no person involved.

Key points

  • Google says attackers have used LLM frameworks to automate vulnerability discovery and exploitation with no human in the loop.
  • The exploit_agent server drafted exploits from patch changes and ran them automatically.
  • Zero-day exploits averaged eight a month in 2025, and reached 22 in August 2026.

Google’s threat intelligence team says it has, for the first time, caught attackers using AI agent frameworks to find and exploit software flaws with no person in the loop. The finding is one of the main points of a report the team published on 16 November, and it describes attacks that run at machine speed, faster than defenders can respond by hand.

The attackers are moving past typing prompts by hand and into multi-agent tools. The team names four open source frameworks it has seen in use: ARTEX, CAIRN, CyberStrikeAI and PentAGI. Google has seen CAIRN, a framework whose agents share one workspace and loop through observe, orient, decide and act steps, in attacks on US aerospace targets.

The most serious case is a server Google calls exploit_agent. In its AI Threat Tracker report, the team says the exposed staging server ran a multi-agent system that checks the old and new code around each patch announcement for popular web software, drafts exploits and runs them without a human. Investigators also found scripts for internal reconnaissance, brute-forcing and planting web shells, as well as removing other attackers’ shells, plus data taken from victims. Those logs hint that French media, entertainment and education sites were in the operator’s sights.

The counts show the pace picking up. By August 2026, the number of disclosed flaws and exploits was close to twice its January level. Zero-day exploits, meaning flaws with no patch yet, grew far more slowly: about eight a month on average in 2025, 11 a month this year, then 22 in August.

Google also says the tools trade stealth for speed. Fully autonomous runs generate traffic that investigators can trace, and they leave distinct traces in the code, but the report says the tools still reach their goals. It expects cybercriminals, who care less about being caught, to keep adopting them.

Chart of vulnerabilities exploited each month, split into n-day and zero-day, from January 2025 to August 2026
Google’s count of exploited vulnerabilities, January 2025 to August 2026 Image: Google

On the defence side, Google says its Security Operations team has deployed detection rules for ARTEX, and Mandiant has published a hardening guide aimed at the speed and scale these tools give attackers. Google’s own offering in this space, AI Threat Defense, pairs Gemini with Mandiant’s expertise and Wiz’s risk ranking, aiming to fix weak spots before attackers reach them.

The report also explains why patched flaws are an attractive target. An attacker can work from a narrow set of material: the code before and after a fix, the release notes and proof-of-concept code. Hunting for unknown zero-days needs a far wider search. The rush has a cost for attackers too, since many LLM-written exploits posted online claim to work and do not, which leaves defenders extra checking to do.

Google says its newest model, Gemini 4 Carbon, has layered safeguards meant to block requests for working exploits or malware, and that outside partners have stress-tested them. The company also says it limits or disables accounts that misuse its tools.

The report expects these operations to move further toward fully agentic intrusions, with less human involvement and more stealth as the models improve.

TPRO Team

The Tech Prolonged newsroom desk. Stories under this byline cover AI, the web and general tech, and are reviewed by our editors before publishing.