EPISODE 4

AI Attacks Are Here. Can Your Response Keep Up?

ARTEX defense on TENEX Signal: how AI agents speed up attacks, why stolen application credentials are the risk, and what to check first.

Podcast

TENEX Signal

19:50

Executive Summary

ARTEX defense starts with the speed of the attack. The FBI issued a flash advisory on ARTEX, an AI-powered tool attackers use to automate parts of real intrusions. Tom Baker and Tyler Moffitt explain how agents take over the trial-and-error work that once slowed attackers down, and they walk through a Microsoft-reported campaign that abused credentials for applications customers had already authorized. Security leaders should review what each application identity can reach, confirm that logs and controls work, and rehearse response with system owners before an incident.

Impactful Moments

00:00

Welcome and framing

01:30

What ARTEX does

04:30

The incident

06:00

Access that outlives the first application

12:00

Automation and evidence

18:00

How TENEX applies this

Markdown File

# AI Attacks Are Here. Can Your Response Keep Up? **TENEX Signal · Episode 4 · November 16, 2026 · 19:50** ## Speakers - **Tom Baker**, Host. Tom opens the episode and guides the conversation for listeners. - **Tyler Moffitt**, Cybersecurity Strategist. Tyler walks through the ARTEX warning, how attackers use AI agents, and what teams can do when there is less time to respond. ## Where to listen - Apple: https://apple.co/4t1LNEK - Spotify: https://open.spotify.com/episode/15PWAONDBzz2xBQo6AJa5m - Listen in your browser: https://www.buzzsprout.com/2655334/episodes/20134770 - All platforms: https://tenexsignal.buzzsprout.com/2655334/follow ## Executive summary ARTEX defense starts with the speed of the attack. The FBI issued a flash advisory on ARTEX, an AI-powered tool attackers use to automate parts of real intrusions. Tom Baker and Tyler Moffitt explain how agents take over the trial-and-error work that once slowed attackers down, and they walk through a Microsoft-reported campaign that abused credentials for applications customers had already authorized. Security leaders should review what each application identity can reach, confirm that logs and controls work, and rehearse response with system owners before an incident. ## Timeline - **0:00–1:30 | Welcome and framing:** Tom introduces the FBI's ARTEX warning and why security teams need to plan for agent-driven attacks now. - **1:30–3:00 | What ARTEX does:** Tyler describes ARTEX as software that takes an objective, uses tools, and decides what to try next, and how Mandiant saw an agent recover from an error. - **3:00–4:30 | How fast attacks move:** Tyler explains that pauses still happen, but defenders should expect the gaps between steps to shrink. - **4:30–6:00 | The incident:** Tyler walks through the Storm-0940 campaign, the sign-in attempts Microsoft reported, and the Huntress example of an application granting itself global administrator. - **6:00–7:00 | Access that outlives the first application:** Disabling the original application does not remove credentials planted elsewhere. - **7:00–8:00 | Spotting expected access:** Tyler explains why high volume can be normal and why teams must understand what each application is supposed to do. - **8:00–9:00 | Controls that held:** Permission restrictions and endpoint protection blocked a move toward domain controllers. - **9:00–10:30 | Checking your defenses:** Start with what a compromised account can reach, what is exposed to the internet, and whether protections actually work. - **10:30–12:00 | What analysts need to know:** Analysts need to follow activity across systems and test whether they have enough records to answer key questions. - **12:00–13:00 | Automation and evidence:** Tyler says automation can gather records, but people must see the evidence behind a decision. - **13:00–14:30 | Approval to act:** Responders need agreed rules for who can shut down a service and when. - **14:30–15:30 | Verifying the action worked:** Teams must confirm the action went through and that the attacker has no other way back in. - **15:30–17:00 | Planning a first test:** Pick one important application and review its permissions with its owner. - **17:00–18:00 | Running the exercise:** Walk through what the responder would do and what the business would experience. - **18:00–19:30 | How TENEX applies this:** Tyler describes automation that gathers context and keeps humans in control of decisions. - **19:30–19:50 | Close:** Tyler leaves listeners with three questions to take back to their teams. ## Transcript ### 0:00–1:30 | Welcome and framing **TOM BAKER:** Hello and welcome to the TENEX Signal podcast. My name is Tom Baker. Today we are looking at an incident that happened on Friday the 13th. The FBI issued a flash advisory about ARTEX, an AI-powered tool attackers are using to automate parts of real intrusions. With me is Tyler Moffitt. Tyler, how are you doing? **TYLER MOFFITT:** Doing well, doing well. How did it happen with this one? **TOM BAKER:** Now then, what should security teams be taking from this? **TYLER MOFFITT:** This is something we need to plan for, right now, as soon as possible. We have incident reports showing attackers using agents to test access, work through errors, and move on to other parts of the environment. This matters because turning access to one account into something useful takes a lot of work. What can that account reach? What permissions does it have? If something fails, what do you try next? Agents can take on more of that work. While they do it, your team still has to understand what is happening, decide what to do, and get someone authorized to act. That puts heavy pressure on the time between those steps. The reporting also shows defenders interrupting these attacks, so there are useful lessons about what worked and what went wrong. I want to walk through an actual example, because it makes this easier to understand than saying attacks are getting faster. ### 1:30–3:00 | What ARTEX does **TOM BAKER:** Before we get into the incident, what exactly is ARTEX doing? **TYLER MOFFITT:** It is software on GitHub that its authors call an AI Autonomous Penetration Testing System. It can take an objective, use tools, look at the results, and decide what to try next. ARTEX brings that kind of workflow to security testing. Attackers have been using it for malicious work, as you can imagine. A human sets the objective and can steer the activity, and the software handles parts of the work in between. We have had password spraying and scanners that work through huge lists of targets for years. What is new is the ability to work through a problem without an operator fixing every individual failure. Mandiant documented an agent hitting an error because a script was too large. It shortened the script, ran it again, and continued. You can see where this leads for an attacker. Troubleshooting takes effort. If software can handle more of it, an attacker has far less to do personally. ### 3:00–4:30 | How fast attacks move **TOM BAKER:** How quickly is this happening? Is it moving straight from one step to the next? **TYLER MOFFITT:** Not always. Mandiant observed extended pauses between bursts of activity. The operations were not all uninterrupted or fully autonomous. But you cannot build your responses and processes in the hope that an attacker will pause at a convenient time. We expect these gaps to shrink. The practical concern is how much work can happen while your team is still making sense of the first signal or alert. An unsuccessful attempt can still produce something your analysts need to investigate. Several attempts can arrive together, and a successful one can open up even more options. So be careful about judging this only by how sophisticated each technique looks. A familiar technique can still cause a serious problem when it is used against a permission nobody has reviewed or a system nobody is watching closely. ### 4:30–6:00 | The incident **TOM BAKER:** You promised to walk us through an incident. Now is your chance. What does it look like? **TYLER MOFFITT:** Microsoft reported that the group it tracks as Storm-0940 used ARTEX in a campaign involving compromised ConnectWise cloud applications. The attackers abused credentials for applications that customers had already authorized. An application identity is essentially an account that software uses. A backup tool needs access to data so it can do its job. A management tool needs permission to manage something. If an attacker gets those credentials, the existing permissions become a very important part of the story. Microsoft reported attempted sign-ins to more than 2,250 customer environments. That is attempted access, not confirmed breaches. Microsoft also separately reported follow-on activity in multiple environments. Another report from Huntress said that in one environment, an application granted itself global administrator. Within about five minutes, new credentials had been added to two other applications. That five-minute window starts at the administrator role grant. It is not the time from the first attempt to full control. ### 6:00–7:00 | Access that outlives the first application **TOM BAKER:** If you shut off the original application, can they still get back in? **TYLER MOFFITT:** They can, if you have not removed access they established elsewhere. Disabling the original application does not remove credentials planted in other applications. You have dealt with the access you first noticed, but you still have to establish what happened through it and after it. It is easy to get distracted by the tool the attacker used and skip past the authority we gave the software. So ask: what does this integration need to read, what can it change, and can it affect another identity? Those are different questions with different levels of risk. ### 7:00–8:00 | Spotting expected access **TOM BAKER:** How do you spot when the application is supposed to have access? **TYLER MOFFITT:** That is the hard part. Consider a backup application that routinely reads a large amount of data. High volume might be completely normal. An alert saying an application accessed a lot of files does not tell you much. You need to understand what the application normally does, who owns it, and whether the activity fits its purpose. A vendor update can also change that behavior, so someone has to keep that understanding current. Abuse of service accounts and application identities is not new. Security teams have dealt with it for a long time. Agents make it more urgent to understand those permissions and recognize meaningful changes. The question is not only whether an account can sign in. It is what that sign-in allows to happen next. ### 8:00–9:00 | Controls that held **TOM BAKER:** So what actually worked against these attacks? Were teams able to stop them with the controls they already had? **TYLER MOFFITT:** Yes. Mandiant reported a separate sequence where an attacker tried to move on to domain controllers, a very common target, since they run the critical identity functions. The attacker ran into permission restrictions. Endpoint protection blocked execution, and defenders cut off access to the server. Those controls made a difference. That does not mean they will catch every attempt, but they are still useful when an agent is doing the work. ### 9:00–10:30 | Checking your defenses **TOM BAKER:** Where would you start checking your own defenses? **TYLER MOFFITT:** Start with what a compromised account or application could actually reach. Does it need all that access, or is some of it left over from when someone set it up? Then look at what is exposed to the internet. It is easy to focus on the main application and forget an old management interface or a secondary login page. CERT-EU's guidance covers familiar ground here. It says to review that exposure, protect management interfaces, keep systems patched, and monitor activity that uses elevated permissions. Look at how those things connect. A small gap can matter a lot if it gives an attacker access to a more powerful account. Finally, check that the protection you expect is actually working. Is endpoint protection reporting? Can you get the logs? Who can make an urgent change? Answer those questions before you rely on those controls during an incident. ### 10:30–12:00 | What analysts need to know **TOM BAKER:** An alert comes in. What does the analyst need to know before deciding what to do? **TYLER MOFFITT:** Say an application unexpectedly changes a permission. Was that change approved? What did the application do before and after? What can it access now? If the analyst has to find three system owners and ask each one to export logs, the attacker keeps working during that time. You need to follow the activity across those systems. A network sensor can explain the connections. Identity and cloud records can explain who signed in and what changed. Each one gives you part of the picture. **TOM BAKER:** How do you know whether you have enough information? **TYLER MOFFITT:** Pick a question you would actually need to answer. For example, could we tell who added a credential to this application, when they added it, and whether they were supposed to do that? Then try to answer it with the records you have. Maybe logging was not enabled. Maybe the record exists but the responder cannot access it. Maybe it was deleted before anyone looked for it. That is something specific you can fix. You do not have to start by collecting every possible event forever. Start with the important systems and the questions you need to answer if something goes wrong. ### 12:00–13:00 | Automation and evidence **TOM BAKER:** I assume parts of that can be automated. **TYLER MOFFITT:** Absolutely. Pulling related records, putting them in order, and checking what we already know about the account or application is useful work that can come off someone's plate. But the person making the decision needs to see the evidence. If the system says something is malicious, show me why. Show me what is still uncertain. A burst of activity is worth investigating, but by itself it does not tell you to shut down a business service. The goal is to give the responder enough context to decide while there is still time to act. You also need to know whether the process holds up when several investigations arrive at once. ### 13:00–14:30 | Approval to act **TOM BAKER:** Here is a scenario many listeners will relate to. What happens if the analyst figures it out but still needs permission to shut something off? **TYLER MOFFITT:** That is a real part of response time. You might be talking about an application that runs backups or supports customer transactions. Disabling it could interrupt something the business truly depends on. You need to work through that with the owner before an incident. What can the responder do? What evidence do they need? Which actions can run automatically, and which need a human to make that judgment call? And who makes the call if the usual owner is not available? It does not help much at two in the morning if nobody knows who is on call. ### 14:30–15:30 | Verifying the action worked **TOM BAKER:** Once they have acted, how do they know it worked? **TYLER MOFFITT:** They have to check. Deciding that something is malicious and stopping the attacker are two different things. Did the action actually go through? Did it affect the [unclear] count or system? Does the attacker still have another way in? Think back to those credentials planted in other applications. Shutting off the first application would not necessarily remove them. Look for the access the attacker may have left behind. That is also why I would look beyond how quickly a team reaches a verdict. How long did it take to restrict the access? And how do they confirm the attacker could not keep going? ### 15:30–17:00 | Planning a first test **TOM BAKER:** So it is all about testing. What is a manageable place for the testing to start? **TYLER MOFFITT:** Pick one important application, such as a backup integration or a management service with broad access. Get its owner together with the people who would investigate and respond. Look first at its actual permissions. What can it reach? What does it still need? Is there anyone who can explain them? Then walk through a dry run. Suppose the application's account gets a new admin permission. Where would that show up? Who would notice? What would it look like? How would they tell whether it was planned maintenance? You can run this as a tabletop discussion, but you want to test the controls directly, so agree on a safe, authorized test. ### 17:00–18:00 | Running the exercise **TOM BAKER:** And carry that through to what they would actually do. **TYLER MOFFITT:** Yes. Have the responder explain what they would need to know and what they are allowed to do. Have the owner explain what would actually happen to the business if they took that action. Then ask how they would check that it worked, including whether the attacker had created another way back in. You may get stuck somewhere in that conversation. Nobody has the right logs, the owner is unclear, or only one person can approve the action. Write that gap down, give someone responsibility for fixing it, and run the exercise again. Practice makes perfect, and now you have something concrete to improve. As a CISO, you also have a clearer reason for where to spend. If the team cannot get the evidence or address the gap, and they are waiting for permission or another alert source, that will not fix the delay. ### 18:00–19:30 | How TENEX applies this **TOM BAKER:** Good stuff. Out of curiosity, how are you applying this at TENEX? **TYLER MOFFITT:** It shapes how we are building the whole service. We want automation to gather context, connect related activity, and work through repeatable investigations, so a human is not starting from scratch every time. We also want people to see how a conclusion was reached. They should be able to inspect the evidence, question it, and take over when human judgment is needed. Response has to work in the customer's environment. What data can we access? Which controls can we use? What have we agreed we are allowed to do? We work through that as part of setting up the service and its scope. What I care about is whether the team can understand what is happening, take an informed action, and then verify that it worked. I would leave you with three questions to take back to your team. What can a compromised identity actually reach? Would we even see the next step? And who can interrupt, and how quickly? ### 19:30–19:50 | Close **TOM BAKER:** Three W's and a bit of an H at the end. Tyler, fantastic. Thank you so much. **TYLER MOFFITT:** Thanks for having me. ## FAQ ### What is ARTEX, and why does the FBI's warning matter? ARTEX is software on GitHub that its authors describe as an AI autonomous penetration testing system. It takes an objective, uses tools, reviews the results, and decides what to try next. Attackers have been using it for malicious work, which is why the FBI issued a flash advisory. ### How do AI agents change the pace of an attack? Agents take on troubleshooting that once needed a human operator. Mandiant documented an agent that shortened a script that was too large and ran it again. Pauses still happen, so plan for the gaps between steps to shrink rather than wait for attackers to stop. ### What did the Microsoft-reported campaign look like? Storm-0940 abused credentials for applications customers had already authorized, and Microsoft reported attempted sign-ins to more than 2,250 customer environments. Huntress reported one application granting itself global administrator, with new credentials added to two other applications within about five minutes. ### If we disable the compromised application, are we safe? Not necessarily. Disabling the original application does not remove credentials planted in other applications. Look for the access the attacker established elsewhere, and confirm they could not keep going. ### How do you defend against ARTEX in your own environment? Start with what a compromised account or application can reach, and whether some of that access is left over from setup. Then review what is exposed to the internet, including old management interfaces. Confirm that endpoint protection is reporting, that logs are available, and that someone can make an urgent change. ## Sources - [Episode blog post: Defending against ARTEX AI attacks](https://tenex.ai/blog/defending-against-artex-ai-attacks) - [Episode 4 on Buzzsprout](https://www.buzzsprout.com/2655334/episodes/20134770) Episode page: https://tenex.ai/podcast/ai-attacks-are-here-can-your-response-keep-up

signal-ep4-ai-attacks-are-here.md

Download the full transcript and technical notes.

FAQ

Frequently Asked Questions

ARTEX is software on GitHub that its authors describe as an AI autonomous penetration testing system. It takes an objective, uses tools, reviews the results, and decides what to try next. Attackers have been using it for malicious work, which is why the FBI issued a flash advisory.

Agents take on troubleshooting that once needed a human operator. Mandiant documented an agent that shortened a script that was too large and ran it again. Pauses still happen, so plan for the gaps between steps to shrink rather than wait for attackers to stop.

Storm-0940 abused credentials for applications customers had already authorized, and Microsoft reported attempted sign-ins to more than 2,250 customer environments. Huntress reported one application granting itself global administrator, with new credentials added to two other applications within about five minutes.

Not necessarily. Disabling the original application does not remove credentials planted in other applications. Look for the access the attacker established elsewhere, and confirm they could not keep going.

Start with what a compromised account or application can reach, and whether some of that access is left over from setup. Then review what is exposed to the internet, including old management interfaces. Confirm that endpoint protection is reporting, that logs are available, and that someone can make an urgent change.