Skip to content
Neural Daily

ResearchNews

GTIG: an AI framework wrote exploits with no human

🇭🇺 Olvasd el magyarul →

Google Threat Intelligence Group has seen, for the first time, an attack framework that runs vulnerability discovery and exploitation with no human in the loop. exploit_agent diffs code before and after patches, writes exploits and runs them. The quarterly report also tracks the spread of LLMJacking.

· 3 min read

A Google Cloud Threat Intelligence blog borítója narancssárga csíkokkal és Threat Intelligence felirattal
Image: Google · Press kit · Source

What did Google find?

According to the quarterly AI Threat Tracker from GTIG and Mandiant, the past quarter brought the first observed case of threat actors using LLM frameworks to fully automate vulnerability discovery and exploitation (Google Cloud). Two months earlier, the previous edition had not seen such pipelines used against live targets (SiliconANGLE).

exploit_agent sat on an exposed staging server, wired to several frontier and local LLMs. It continuously diffs popular web application code before and after patches, looks for vulnerabilities, drafts exploits and executes them automatically. The server also held data stolen from victims and post-compromise scripts for reconnaissance, brute-forcing, installing webshells and even removing other attackers' webshells.

A Russian comment in the config file read "LLM API keys for auto-exploitation (exploit_agent pilot)". Telemetry suggests the operator targeted French media, entertainment and education websites. The report names no victims, CVEs or actor.

Why n-days?

GTIG argues that pointing LLMs at already patched n-day bugs is more efficient, because the source material is finite: two code versions, the release notes and related PoC code. Hunting zero-days needs a much wider net. The practical result: the gap between a patch and a weaponized exploit shrinks sharply.

Oszlopdiagram: kihasznált sérülékenységek havi száma n-day és zero-day bontásban, 2025 januártól 2026 augusztusig
Google · Screenshot · Source

Source: Google Threat Intelligence Group

The numbers point the same way. Between January and August 2026, disclosed vulnerabilities and exploits nearly doubled. Zero-day exploits rose from a 2025 average of 8 per month to 11, then spiked to 22 in August.

What else is in the report?

  • Open-source multi-agent pentest frameworks in live intrusions: ARTEX, CAIRN, CyberStrikeAI, PentAGI. CAIRN was used against US aerospace targets. More on CAIRN:
  • [object Object]
  • .
  • LLMJacking: stolen cloud credentials used to burn other people's LLM quota. Single-day fraudulent charges exceeded $100,000 in some cases, and reached $1 million in total in the worst ones.
  • LLM Mixers: illicit storefronts reselling model access through stolen keys.
  • Supply chain: espionage groups tampering with coding assistant configurations to drop backdoors during local execution.

Google says these tools currently trade stealth for speed: they leave plenty of forensic traces, which makes them detectable. Mandiant published a hardening guide, and Google SecOps shipped rulepacks to detect ARTEX.

What it means for operators

The patch window keeps shrinking, especially for internet-facing web apps. If a CMS or plugin update waits days or weeks, you are now racing a machine. Treat LLM API keys like cloud admin rights: spending alerts, key rotation, IP restrictions. A related result: GLM 5.3 built Firefox exploits from the patch gap.

Key takeaways

  • GTIG saw a fully autonomous exploit pipeline in the wild for the first time.
  • exploit_agent writes and runs exploits from patch diffs, no human involved.
  • Zero-day exploits: 11 per month instead of 8, 22 in August.
  • LLMJacking: fraudulent charges above $100K per day.

    By Mikola Máté Patrik

    Produced with the help of AI, reviewed by an editor. How we make our articles

    Reviewed and edited by the Neural Daily desk. About

    Related articles