On November 13, the FBI released a FLASH alert on ARTEX, an open-source framework that uses AI agents to carry out intrusions with minimal human input. Actors have used it since at least August 2026 against financial services, government, healthcare, energy, telecommunications, and other sectors. Some intrusions went from initial targeting to privileged access in under three hours (FBI).
Three days later, Google Threat Intelligence Group (GTIG) published its own report on agentic attack frameworks, with ARTEX among them.
The three-hour number got the headlines. The line I keep coming back to is further down the alert. During internal reconnaissance, intrusions using ARTEX “inventoried endpoint sensor coverage to identify Windows machines without CrowdStrike Falcon.”
That’s a coverage gap analysis. It ran inside the victim’s network, as one step in an intrusion driven by an AI agent working from a goal someone described in natural language.
Mapping sensor coverage isn’t new. Red teams have done it for as long as there have been sensors. What caught my attention is the order of operations. The agent treated the defender’s coverage map as recon data. If an agent builds that map mid-intrusion, we should have ours built before it shows up.
Last month, I wrote about a Unit 42 report on an AI-assisted espionage campaign in Same Tricks, Faster Attacks: The Detection Engineering Problem AI Doesn’t Solve, and shared some of the same thinking on LinkedIn. The attackers copied /bin/bash, renamed it, and used it to run commands as root. The behaviors were old. The speed of the variations was new. ARTEX applies the same pattern to an entire intrusion.
GTIG describes attackers moving past prompting models by hand to running multi-agent frameworks that carry out multi-stage intrusions. Alongside ARTEX, it has seen CAIRN and PentAGI used in the wild, plus experimentation with CyberStrikeAI and HexStrike AI.
GTIG is also clear the underlying mechanics haven’t changed. Compromised credentials and unpatched systems still do most of the work. Speed and scale are what’s different.
The FBI’s ATT&CK mapping of ARTEX activity reads like a greatest-hits list. Valid Accounts and Trusted Relationship for initial access. Web shells for persistence. Kerberoasting and DCSync for privilege escalation. Pass-the-Ticket, Pass-the-Hash, WinRM, and Kerberos relay attempted against domain controllers for lateral movement.
Every one of these has been in ATT&CK for years. We’ve emulated them, written about them, and built detections for them.
What’s new is the loop. CAIRN, for example, runs an Observe, Orient, Decide, Act cycle. Specialized agents share findings, try a technique, read the result, and adjust, starting from as little as a target IP address.
That loop changes what “covered” means. When a human operator gets blocked, there’s usually a pause to rethink. An agent in a feedback loop moves to the next option. So the useful question isn’t whether you cover a technique. It’s whether you cover the alternatives an attacker reaches for when the first one fails.
These frameworks are noisy. GTIG says they currently “trade stealth for velocity,” leaving forensic traffic, recognizable identifiers, and code-level traces behind. The FBI agrees ARTEX generates enough noisy traffic that it “may trigger alerts,” but warns its speed “provides limited time for coordinated defender response before objective completion.”
So detection isn’t the hard part right now. Time is.
Think about the usual cycle. A threat report lands. An analyst maps the techniques. Another team checks whether existing controls cover them. If there’s a gap, someone builds, tunes, and deploys a fix. In most organizations I talk to, that takes days or weeks.
That cycle was built for adversaries who also needed days.
Adding AI to the same cycle doesn’t change the math. In my previous post I called this the Goldilocks problem. Make a detection too specific and a rename beats you. Make it too broad and your analysts learn to ignore it. Point AI at a report with no behavior context and it writes a rule for the exact command line it was shown. Faster at the wrong thing is still the wrong thing.
One detail in the FBI alert deserves more attention than it’s getting. When ARTEX attempted Pass-the-Hash against domain controllers, endpoint protection blocked it.
Nobody wrote a rule mid-intrusion. The control was in place and configured for that behavior before the attacker arrived.
The alert doesn’t report outcomes for the other three lateral movement attempts. That’s worth noticing. One blocked technique out of four tried is a good result for one control. It’s also a reminder of the loop above. The agent had three more options queued.
There’s another reason behavior-based defense holds up here. ARTEX doesn’t ship its own model. It plugs into whichever LLM provider the operator picks, and the FBI says attacker instances mostly ran DeepSeek models. The model changes. The repo gets forked. The behaviors stay the same. Defenses tied to behavior don’t care which model is driving.
GTIG also found a framework it calls exploit_agent on an exposed staging server. It continuously compares popular web application software before and after patch announcements, drafts exploits for what changed, and runs them without a human in the loop. GTIG says this is the first time it has seen threat actors use LLM frameworks to fully automate vulnerability discovery and exploitation.
Note what it targets. Patch diffing works on flaws that already have a fix. Patching still works. The time you have to do it is what’s shrinking.
The volume is climbing too. GTIG says vulnerability disclosures and exploits nearly doubled between January and August 2026. Zero-day exploits averaged 8 a month in 2025 and 11 a month in 2026, then hit 22 in August.
One SecOps leader told us his team can’t keep pace with AI-discovered vulnerabilities anymore, so they’re shifting focus to where they can interrupt the attack chain. I think that’s the right instinct. If you can’t patch everything first, you need to know whether you can stop what comes after the exploit.
GTIG expects these operations to need less human involvement and to get stealthier over time. ARTEX isn’t going away either. Its developer made the GitHub repository unavailable on October 8, but forks, mirrors, archived copies, and a commercial version called ARTEX PRO are still circulating. Researchers found more than 1,600 publicly exposed ARTEX consoles as of early November.
Right now these attacks are fast and easy to see. That makes this a good time to test your defenses against them. The FBI handed us an ATT&CK mapping. Turn it into an emulation plan, run it with the adversary emulation tools you already use, and see what fires.
Each item comes from what the FBI observed ARTEX doing, or from its recommended mitigations:
Every report like this raises the same questions for a specific environment. Which of these behaviors apply to us? Do our controls cover them, and are they configured to? Where are the gaps? Which change closes the most of them at once?
The ARTEX reports didn’t change those questions. They shortened the deadline for answering them.
That’s where I think AI helps defenders most: shrinking the time between “here’s how this adversary operates” and “here’s what we changed.”
It’s also the work we focus on at Tidal Cyber under Threat-Led Defense. We connect procedure-level adversary intelligence to the defensive tools a team already owns, so it can see where the gaps are and decide what to fix first.
The ARTEX agent started by asking where the sensors weren’t. Good question. Ask it before they do.