Threat intel / Techniques and trends
Attackers run open-source agentic pentesting frameworks in live intrusions
Google says several threat actors now run open-source AI pentesting frameworks such as ARTEX, CAIRN and PentAGI in live intrusions, and one framework exploited a flaw with no human in the loop.
High severity. Active. Moderate confidence. First reported 30 Oct 2026, last activity 16 Nov 2026. Briefing updated 18 Nov 2026.
AI-written summaries of public reporting, each checked against at least two sources from different publishers. Check the sources before you act. Last refreshed 2026-11-18 14:09 UTC. How this is made.
What is not settled: The November 16 GTIG post was read through a summarizing fetch, not line by line. The October 30 GTIG post and the BleepingComputer article were seen only as search excerpts, and they agree on the framework names and on the agentic abuse. Sources disagree on HexStrike AI and CyberStrikeAI. One account says GTIG observed only CAIRN and PentAGI in real attacks, and another lists all five names as abused. This briefing says Google names the five tools and does not claim every one was seen in an intrusion. The FBI warning and the Vapor Panda figures for ARTEX (200 organizations) come from secondary coverage, so they are left out. The CVEs are as Google names them. CISA's KEV catalog could not be checked from this environment, so no kev value is recorded for any of them.
In plain words
What happened
Attackers have started using open-source AI tools that plan and carry out hacking steps on their own. Google saw several groups use them in real break-ins, and it found one tool that scanned for a flaw, wrote an exploit and ran it without a person in the loop.
Who is behind it
Google names several groups, including Iranian and North Korean-linked actors and a China-linked group. It does not give a single actor for the framework use.
What they want
Google does not state one goal. The tools mainly add speed and scale to the same intrusions attackers already run.
What they did
Google saw CAIRN used against a US aerospace target and a customized PentAGI-based platform scan organizations in Asia from late August to early October 2026. It found an exposed server running a framework that compares patches, drafts exploits and runs them. Its telemetry points at French media, entertainment and education websites.
Who is at risk
Any organization exposed to the internet, since the tools scan and exploit faster than a human team can respond. Google's examples include media, education, aerospace and Asian targets.
What to do
- Patch internet-facing software quickly, starting with the flaws named in Google's post.
- Watch for fast, repeated scanning and exploit attempts across many hosts at once, which is what automated frameworks produce.
- Check logs for unexpected AI agent or automation tools on servers and developer machines.
- Shorten the time between a patch and its deployment on exposed systems.
The technical version, for hunting and patching
What happened
Google's November 16 GTIG AI Threat Tracker reports that attackers have moved from manual prompting to multi-agent offensive frameworks. Google names ARTEX, CAIRN, CyberStrikeAI, PentAGI and HexStrike AI. Google observed multiple threat actors using open-source agentic pentesting frameworks during live intrusions, and it describes an exposed staging server running a framework that does continuous differential patch analysis, drafts exploits and executes them.
Who is behind it
Several groups. Google names CALANQUE ION (Iranian, previously APT42), MIDNIGHT NEPTUNE (DPRK-linked) and LEAKY CASTLE (a PRC contractor), and says a PRC-nexus group was building its own automated framework. Google does not attribute the framework intrusions to one actor.
What they want
Google describes espionage and financially motivated intrusions, and says autonomous execution trades stealth for speed. It also reports LLM-backed monetization, including LLMjacking losses of up to $1 million in the most severe case.
What they did
Observed activity includes CAIRN in an intrusion against a US aerospace target, a PentAGI-based platform scanning Asia-based organizations, and the PaperCut chain (CVE-2026-81578 and CVE-2026-82078), which GreyNoise reported being exploited four days after disclosure. Google says the exploit agent found the first fully automated discovery and exploitation without a human in the loop. It also notes that LLM-generated proof-of-concept code for PHP flaws was mostly non-functional.
Who is at risk
Internet-facing software with known or newly disclosed flaws, and developer and cloud environments that hold credentials. Google reports that a Q2 2026 case involved a cloud compromise and an autonomous multi-agent framework.
What to do
- Treat the framework names in Google's post as indicators to hunt for on exposed hosts and developer systems, and review them before blocking.
- Prioritize patching for chained flaws such as the PaperCut pair and the PHP issues named in the post.
- Restrict and rotate secrets stored in code repositories, CI systems and AI tool configuration.
- Review Google's hardening guidance and the detection rulepacks it published for ARTEX before relying on them.
Vulnerabilities
| CVE | Product | CVSS | CISA KEV | Fixed in |
|---|---|---|---|---|
| CVE-2026-81578 | PaperCut | Not checked | ||
| CVE-2026-82078 | PaperCut | Not checked | ||
| CVE-2026-109618 | PHP | Not checked | ||
| CVE-2026-109619 | PHP | Not checked | ||
| CVE-2026-40217 | LiteLLM | Not checked |
Sectors: Manufacturing, Media, Education. Regions: Global.
Open next
Sources
- GTIG AI Threat Tracker: Agentic Threat Activity Crosses a Significant Threshold, Google Threat Intelligence Group, 2026-11-16 (vendor)
- Artex and friends enter agentic offensive frameworks, Google Threat Intelligence Group (vendor)
- Google warns that fully-automated AI attacks are coming soon, BleepingComputer (news)
Revision history
- 2026-11-18: First published.
Older: SCKit abuses ScreenConnect to reach MSP customer machines.