An exploit pipeline with nobody at the keyboard
Google Threat Intelligence Group said on Monday that it has seen threat actors use large language model frameworks to automate the discovery and exploitation of software vulnerabilities “without a human in the loop for the first time”. The finding is the centrepiece of the group’s quarterly AI Threat Tracker, which covers incidents from August to October 2026.
The clearest example came from an exposed staging server GTIG found hosting a multi-agent framework its operators called exploit_agent. According to the report, it runs continuous differential analysis of popular web applications — comparing the code before and after a patch, working out what was fixed, drafting an exploit for it and running it. Other scripts on the same server handled internal reconnaissance, brute-forcing and installing webshells, and one removed webshells left by rival intruders. Victim telemetry pointed at French media, entertainment and education websites, and an environment file on the server carried comments in Russian.

The frameworks are off the shelf
Most of the tooling GTIG describes is public. It names ARTEX, an agentic penetration-testing tool it says is being actively abused and for which Google SecOps has shipped detection rulepacks, and CAIRN, a multi-agent framework built around a shared “blackboard”, which GTIG says was used in intrusions against targets including the US aerospace sector. A customised build of PentAGI was used from late August to early October 2026 for reconnaissance and vulnerability scanning against organisations in Asia. CyberStrikeAI and HexStrike AI, which exposes more than 150 offensive tools through a dozen sub-agents, were seen in experimentation rather than live operations.
There is a catch, and GTIG is explicit about it: these tools “currently trade stealth for velocity”. Autonomous execution is noisy, leaves recognisable identifiers and produces forensic traces. The group expects that to change, writing that as the tools and the underlying models improve it anticipates “not only reduced human involvement in these operations but also increased stealth and sophistication of the activity”.
Two cases, and a flood of fake exploits
The report attaches numbers to the acceleration. Zero-days, GTIG says, averaged eight a month in 2025 and eleven a month in 2026, then spiked to 22 in August 2026, with disclosures and exploits close to doubling between January and August.
In late August, two PaperCut zero-days — CVE-2026-81578, an authentication bypass, and CVE-2026-82078, unsafe dynamic class loading — were chained for unauthenticated remote code execution. Working exploits appeared within a day of disclosure, and GreyNoise reported a campaign four days later, likely LLM-assisted, against hundreds of organisations; in it, GTIG says, the attacker went from first foothold to domain administrator within minutes. After two PHP flaws were patched on 21 October, hundreds of claimed proof-of-concept exploits appeared that did not work, most of them apparently model-generated, adding verification work for defenders rather than risk.

Stolen keys, resold by the million tokens
The other growth area is LLMjacking. Mandiant’s incident response saw single-day fraudulent charges on hijacked AI accounts that sometimes exceeded $100,000, and worst cases where cumulative abuse reached about $1 million before it was stopped. One botnet or proxy service generated millions of API calls in roughly 24 hours. Attackers scrape credentials from public repositories, hijack sessions and exploit CVE-2026-40217, a LiteLLM sandbox escape that turns a limited virtual key into an administrative one. The stolen access is resold through “LLM mixers”, storefronts on Russian- and Chinese-language markets charging between $0.10 and $25 per million tokens.
GTIG also reports that influence-operations actors have moved from experimenting with models to building them into their pipelines, including one commercial platform designed to run more than 100,000 accounts.
What to watch
Google says it disabled the assets tied to the activity and that DeepMind used the findings to tighten classifiers, and it is pitching its own answer, Google AI Threat Defense, which bundles Gemini with Wiz, CodeMender and Mandiant. The claim to test over the next quarter is GTIG’s own forecast: that cybercrime groups will keep adopting noisy autonomous tooling because reach compensates for a lower success rate, while espionage actors use it selectively.