ARTEX and Autonomous Offensive Frameworks
Published 11/18/2026
They’re here. Time’s up. The window is closing.
I despise hyping things up, but now that the FBI is issuing a flash warning about autonomous attacks, the window is down to just a little crack of an opening left for us to get our defenses in place.
On November 17 we hosted an emergency CSA CISO Huddle with Gadi Evron on ARTEX and autonomous attacks. We pulled it together with less than a week’s notice, and roughly 600 people joined. Our speakers and attendees all shared what they are seeing, and there is a massive increase in the use of Autonomous Offensive Frameworks in active campaigns, by a wide range of threat actors. These tools, like ARTEX, aren’t magic destruction boxes, they just take the capabilities of a decent pen tester, automate the process, and then operate relatively quickly.
A quick catch-up for anyone who hasn’t been buried in the reporting: ARTEX is an open source, LLM-driven penetration testing framework that appeared on GitHub in late July. You point it at a target, wire it to whichever commercial or open models you like, and agents plan and run the attack. The author took it closed source on October 8 after it started showing up in real intrusions, but the open versions had already spread. Rami McCarthy has a good write-up on how ARTEX is now a family of forks and backdoored copies and should be treated that way, and Censys counted 1,465 exposed ARTEX consoles on the Internet in early November. (For the record, the attackers are also bad at cloud security.)
The FBI issued a flash advisory saying some ARTEX intrusions went from targeting to active compromise in under three hours. CrowdStrike tracked VAPOR PANDA running a single ARTEX controller against more than 200 organizations in at least 14 countries over 11 days, with a human stepping in mainly at decision points. The operator even told the agents to go find hosts that didn’t have a CrowdStrike sensor. And Barracuda disclosed that an ARTEX-driven intruder came in through a valid cloud access key belonging to a third-party provider it had stopped working with and never revoked, then touched about 10 cloud accounts in roughly 44 hours.
Read those reports back to back and very little of what they describe needed a zero-day. Amazon’s researchers, quoted in SecurityWeek’s coverage, said the attackers succeeded where “basic security hygiene was missing.” The problem is the basics are hard once we have to scale them.
Datadog's analysis of ARTEX in AWS is a great article to start with. The agents got in with leaked long-lived access keys, or by self-registering accounts in Cognito user pools that allowed anonymous sign-up. Once inside, the most common thing they did was call FilterLogEvents and search CloudWatch logs for terms like “Invoice” and “Billing,” and in later sessions for presigned S3 URLs, internal IP addresses, and AKIA access keys. Think about that for a minute. A human attacker rarely has time to read every log group in an account. An agent will, for as long as the token budget holds out, and any secret a developer ever dumped into a debug statement becomes something it can use.
The same report has encouraging details. In one Cognito campaign the agent registered accounts and then ran a long list of discovery calls (ListBuckets, ListSecrets, DescribeLogGroups, and more). Every one failed, because the identity pool’s authenticated role didn’t grant those permissions. The agent kept iterating anyway and got nowhere. A boring least-privilege role did its job, and the agent was really noisy and easy to detect.
Agentic attacks reward neglect. The forgotten access key, the vendor integration nobody turned off, the open sign-up flow, the application log with a password in it. An attacker that never gets bored will find them eventually, and these tools let it look across hundreds of organizations at once. BUT THIS DOESN’T MEAN YOU ARE NEGLIGENT FOR HAVING THESE HOLES! Defense is hard. Consistency is harder. Consistency at scale is borderline impossible.
So where do you start? Our Frontier-Ready advice doesn’t really change, but this wave does highlight some areas to consider moving first on. If you have a red team (or contract with one), have them use an AI-driven framework and run a test (be careful, without the right controls they may break things, so make sure the operator knows what they are doing). Worst case, point your cyber-capable model internally with just read access.
Second, get moving on deception technology. Honey tokens/networks/etc. are highly effective for early detection. Third is hardening—basically what I recommended last month on identifying critical assets and walling them off. And no, this isn’t easy.
The Frontier Ready Maturity Model goes much deeper on all of this, and the public version is free. Last month I said don’t panic, but stop treating this as next year’s problem. I’d now call it this week’s problem, though most of the fixes are things your team already knows how to do.
Unlock Cloud Security Insights
Subscribe to our newsletter for the latest expert trends and updates
Related Articles:
DLP and UEBA Are Asking the Wrong Insider Risk Questions
Published: 11/18/2026
The AI Threat Landscape in 2026: What You Need to Know
Published: 11/18/2026
CSA STAR for AI Certifications: What are They?
Published: 11/17/2026
AI Governance Is Cloud Governance: From Posture to Proof
Published: 11/17/2026
.jpg)


.jpg)

.png)
.jpeg)
.jpeg)

.jpeg)