November 18, 2026

From First Experiments to Exploits With No Human at the Keyboard: A Timeline of How Attackers Handed Their Work to AI Agents

Google’s threat intelligence team and Mandiant say attackers now let AI agent frameworks run whole stages of real intrusions, and one operation turned freshly published patches into working attacks without a person in the loop. Here is how 2026 got there, step by step, and what it means for patching at a small business.

A security analyst in a gray hoodie typing at a keyboard in an operations room lined with monitors and a large world map display

Google Threat Intelligence Group (GTIG) and Mandiant have published new research on how criminals and state-backed hackers are using artificial intelligence, and the headline finding is a first: an attack operation that compared software before and after a security patch, worked out the flaw, wrote an exploit and launched it, all without a human stepping in. Help Net Security summarized the report, and The Hacker News added detail on patch analysis and the tools involved. Read in order, the findings show how quickly this shifted during 2026.

How the year unfolded

January to August 2026: more bugs, more exploits

GTIG’s September analysis of vulnerability data found that disclosures and exploits nearly doubled over these months. Zero-day exploits reached 22 in August, against a 2025 average of 8 a month.

Since at least August: agent frameworks in live intrusions

Attackers began using ARTEX, an open-source agentic penetration-testing framework, in real attacks, according to the FBI. GTIG also saw CAIRN, CyberStrikeAI, PentAGI and HexStrike AI in use, plus homemade harnesses. These tools can find and exploit simple flaws and adjust their approach based on how a target responds. Users include financially motivated criminals and suspected China-linked espionage groups, and GTIG found one PRC-nexus group apparently building its own automated framework.

August: AI-run influence campaigns

GTIG investigated a platform designed to automate entire influence campaigns about regional politics in Europe and the Middle East. Its prompts asked models to imitate local slang and add deliberate typos, and images were roughened to pass as smartphone photos.

September: patches reverse-engineered by machine

The Hacker News reports that GTIG observed attackers using language models to compare code versions and quickly weaponize flaws in PaperCut NG/MF. Google notes this is attractive to attackers because they only need to study a limited set of material: code changes, release notes and published research about a specific fix.

October 21: a flood of fake exploits

After patches for two PHP flaws (CVE-2026-109618 and CVE-2026-109619) shipped, hundreds of AI-generated proofs of concept appeared online that claimed to work but did not, and defenders had to check each one.

November 13: the FBI weighs in

The bureau issued a FLASH alert on ARTEX, warning of its use since at least August against finance, government, technology, defense, healthcare, energy and other sectors.

November 16: Google publishes the “no human in the loop” finding

GTIG says it saw this for the first time in the past quarter. The evidence came from an exposed staging server running a multi-agent framework called exploit_agent, wired to several commercial and locally run models. It compared popular web applications before and after patch announcements, located the flaw, drafted an exploit and ran it. The server also held stolen data and scripts for brute-forcing passwords and for installing web shells and removing ones left by rival attackers. A configuration comment, written in Russian, labeled the model API keys as being for an “exploit_agent pilot.” Victim data suggested a focus on French media, entertainment and education websites.

Other findings in the same report

  • Stolen AI access for sale: Mandiant handled several 2026 “LLMjacking” cases in which attackers used a victim’s cloud accounts to run AI workloads. Charges topped $100,000 in a single day in some cases and reached $1 million in the worst before being stopped. The access is resold through underground “LLM mixers” at $0.10 to $25 per million tokens.
  • North Korean operators: Fake IT workers use AI to write resumes and sound convincing in interviews. A cluster tracked as MIDNIGHT NEPTUNE used AI-generated recruiter photos and fake meeting sites with ClickFix lures, then edited coding-assistant settings files in company repositories so the tool would download a backdoor.
  • Phishing kits with a chatbot: A Chinese-language phishing platform, Jieyou, lets customers customize lures and fake login pages by typing plain-language requests.

The silver lining, for now

GTIG says today’s frameworks “trade stealth for velocity,” leaving noisy traffic and code traces that make them easier to catch. Google expects that to change as the tools and models improve, with less human involvement and quieter activity.

What this means for your company

The time between a vendor publishing a patch and attackers using it is shrinking, because the slow, skilled part of the job is being automated. For a small business, that makes patching speed a real security control: know which internet-facing systems you run (website software, VPN, remote access, file-sharing and print-management servers), and apply critical updates within days, not at the next quarterly window. Remove or protect anything that does not need to face the internet, monitor cloud billing for sudden spikes that could signal stolen access, and treat unexpected recruiter outreach and “run this diagnostic” requests with suspicion. The good news in Google’s findings is that these tools still make noise, so basic logging and alerting will catch many of them.

If you are not certain how long your critical patches take to reach every exposed system, we can help you measure and shorten that window. Reach out to Layer 7 Data Solutions to discuss how we’ve helped companies similar to yours.

About The Author