SANS Stormcast Friday, May 8th, 2026: AI Generated Dashboard; Ivanti Patches; Redis Vuln; @sans_edu Marcio Enriquez

Summary

SANS Stormcast | Friday, May 8th, 2026 — Episode Summary

Podcast: SANS Internet Stormcenter Daily Cyber Security Podcast (Stormcast)
Host: Johannes B. Ullrich
Guest: Marcio Enriquez
Episode Theme:
A fast-paced, expert-driven update focused on key events in cybersecurity: innovative use of AI for dashboards, urgent vulnerability patches (Ivanti, Redis), and an in-depth research segment on the risks of AI-driven automated defensive tools in enterprise environments.

---

Episode Overview

Johannes Ullrich delivers a concise briefing on recent network security developments, touching on the power of AI-generated dashboards for log analysis, urgent patches for Ivanti and Redis, and handing over the floor to SANS Edu student Marcio Enriquez. Marcio shares insights from his recent research into how AI-driven defense tools could be exploited to cause widespread enterprise disruption—a timely topic given the rise of autonomous security solutions.

---

Key Discussion Points and Insights

1. AI-Generated Dashboards for Honeypot Logs

00:15–02:18

• Innovative Approach:
  Undergraduate intern Eric Rolten created a Python script to standardize and summarize honeypot log data, then leveraged Claude AI to generate dynamic dashboards.
• Improvements Over Traditional Dashboards:
  • Problem: Older dashboards are often too static, built for specific past threats.
  • Solution: The AI-generated dashboards are more adaptive, highlighting new patterns, big threat actors, and vulnerabilities being exploited in the present data.
• Host Impression:
  “I'm actually quite impressed with the different dashboards that came up with here...having the simplicity of creating essentially wipe coding these dashboards, I think makes a lot of sense.” — Johannes Ullrich (01:32)

2. Ivanti Vulnerabilities and Patch Advisory

02:19–03:40

• Newly Patched Vulnerability:
  Ivanti's May patch fixes several issues, including one already exploited in the wild—but requires admin credentials to leverage.
• Credential Lifecycle Reminder:
  Credentials potentially compromised before January’s update may still be vulnerable; rotation and patching are critical.
• Host Recommendation:
  1. Apply all new patches.
  2. “Double check that you recently rotated your credentials on your Ivanti Endpoint Manager mobile devices.” — Johannes Ullrich (03:21)
• AI in Vulnerability Discovery:
  Ivanti, alongside the industry, is increasingly using AI to proactively discover vulnerabilities. Unlike some vendors, Ivanti commits to disclosure even if the issues are found internally with AI.
  • “Ivanti is using these tools and not just waiting for some researchers...but instead proactively...find as many vulnerabilities as they can.” — Johannes Ullrich (03:33)

3. Critical Redis Remote Code Execution Vulnerability

03:41–04:30

• Patch Available:
  • Resolves a remote code execution (RCE) flaw in Redis; credentials are required to exploit.
• Host Context:
  Redis databases are often exposed or shared, and credentials are sometimes distributed widely.
• Key Recommendations:
  • Limit direct network access.
  • Enforce the most robust authentication methods possible.
  • Read advisory footnotes for additional setup tips.

4. Research Spotlight: AI-driven Defensive Automation & Exploitation Risks

05:20–14:33

Guest Introduction & Context

05:20–05:36

• Marcio Enriquez: Lead Service Manager in exposure management at a Fortune 500 company, over a decade in cybersecurity.

Origins of the Research Paper — Real-World Incident

05:47–07:03

• Incident:
  A phishing alert led to the discovery that Microsoft’s Automatic Attack Disruption (an autonomous AI tool) had taken sweeping actions—locking users (including managers) out—without the team’s awareness.
• Catalyst Question:
  “One of our VPs had a question, he said, well, could this be used against us? And that question really provoked my thought process and thus the research topic.” — Marcio Enriquez (06:23)

Research Focus and Experiment Design

07:04–09:09

• Research Question:
  Can defensive AI/automation be manipulated to induce enterprise disruption?
• Lab Setup:
  • Realistic enterprise environment: Azure AD, E5 trial license, domains, user personas, virtual machines, segmented attacker machine.
  • Goal: Not system compromise, but “can we hit the triggers that caused the automated actions to take? Can we simulate that to make it force user containment and how far would it go?” — Marcio Enriquez (08:44)

Findings: Narrow Context & Over-Enthusiastic AI Responses

09:09–11:43

• Historical Parallels:
  Past features (Cisco, account lockouts via password spray or mass internal port scans) caused broad disruption from relatively trivial triggers.
• Results:
  • Triggering AI protections (attack disruption) locked out 18 test accounts.
  • Enabling hybrid cloud password writeback locked out even the admin account:
    “It also locked out my domain admin account on the local environment, which was a lot of fun to get back into.” — Marcio Enriquez (10:10)
• Key Limitation of Today’s AI:
  Even advanced AI tools exhibit "artificial narrow intelligence," acting without full situational context.
  • “All the language learning models...are really good at what it's specifically coded or designed to do, meaning contextually it may not fully comprehend the ramifications of its actions.” — Marcio Enriquez (10:41)
  • Example: AI disables domain admin account—something a human would reconsider due to recovery/availability concerns.
• Call to Action:
  Always use guardrails: break-glass, emergency accounts, runbooks.

Configuration, Controls, and Rollback Mechanisms

11:43–13:19

• Exclusion Options:
  • Microsoft added options to exclude certain users/machines from autonomous containment—but it’s not required to configure or visible by default.
  • “If you don't know to look for it, you would never know that it existed.” — Marcio Enriquez (11:55)
• Recovery:
  • No simple “undo” feature; manual recovery via privileged cloud accounts may be required.
• Logging:
  • Autonomous actions are thoroughly logged and can be rolled back if identified in time.

Larger Enterprise Risk

13:19–14:33

• Adversarial Knowledge Issue:
  “If an adversary...knows what modern equipment you're using...they no longer are trying to hide themselves to cause disruption. They're more like, let's purposefully find these high yield triggers to make actions taken to force a containment that maybe a human would not do.” — Marcio Enriquez (13:45)
• Visibility & Prevention:
  • Extensive logging exists, but attackers can trigger well-documented responses for maximum organizational disruption (e.g., locking key admins out at a critical time).

---

Notable Quotes & Memorable Moments

• On Continuous Dashboarding:
  “You may have created them years ago or such for a particular problem you had back then, but they often aren't adjusted very frequently to really sort of point out what's new and what's interesting in the data.” — Johannes Ullrich (00:46)
• On Autonomous Defense Pitfalls:
  “A normal analyst seeing that the domain admin account is doing something malicious, may not so quickly disable that account without first making sure that we have access back into our network...” — Marcio Enriquez (10:53)
• Configuring AI Tools:
  “It's not a requirement to fill in a user to have this enabled. So, you know, if you don't know to look for it, you would never know that it existed.” — Marcio Enriquez (11:53)
• Enterprise Recovery Challenge:
  “There’s no easy undo button...” — Johannes Ullrich (12:40)
  “I had to go in there, log in through [my Azure account] and leverage a live session to undo and disable all the actions that were taken by the autonomous agent.” — Marcio Enriquez (12:47)

---

Important Segment Timestamps

• AI-Driven Dashboard Discussion: 00:15–02:18
• Ivanti Vulnerability Warning: 02:19–03:40
• Redis RCE Advisory: 03:41–04:30
• SANS Edu Interview Start (Marcio Enriquez): 05:20
• Origins of AI Attack Disruption Research: 05:47–07:03
• Research Findings & Exploitation Risks: 09:09–14:33

---

Closing Remarks

This Stormcast episode highlights both innovative progress and the growing complexity in cybersecurity: automation and AI bring immense value, but introduce new attack surfaces—especially if adversaries can trigger overzealous automated responses to cause disruption. Patching, credential hygiene, careful configuration, and strong incident response plans remain crucial.

For more detail or to read Marcio Enriquez’s full research paper, see the show notes link provided in the episode.

Transcript

[A]

Hello and welcome to the Friday, May 8, 2026 edition of the SANS Internet Storm Center's Stormcast. My name is Johannes Ulrich, recording today from Jacksonville, Florida, and this episode is brought to you by the SANS Edu Graduate Certificate Program in Cybersecurity Engineering. In diaries. Today we had another diary by one of our undergraduate interns, in this case Eric Rolten, wrote about how create better sort of analytics UIs to analyze honeypot logs. And of course this can be adopted to other types of logs too. A problem with all these dashboards and such is often that they're too static. You may have created them years ago or such for a particular problem you had back then, but they often aren't sort of adjusted very frequently to really sort of point out what's new and what's interesting in the data. So what Eric did here is Eric wrote a Python script that will first of all summarize all the data and create a summarized version in a standardized format. And then, well, Eric, as the kids do it these days, sent AI after the data and had AI Claude, in this case create a dashboard to display the data. And I'm actually quite impressed with the different dashboards that came up with here. So there are some generic security dashboards that have things that you would sort of expect, things like how many ports scanned, how many requests we had, different time series and the like. But then there are also sort of some things like what are, for example, the big actors and then what are those actors doing. So essentially a dashboard that will just summarize the traffic for a particular actor or for a particular vulnerability that's being exploited. So yes, lots of classic dashboards have these kind of features. But of course, having the simplicity of creating essentially wipe coding these dashboards, I think makes a lot of sense and like I said, kind of impressed by the output that was generated using this particular approach. Well, let's continue with some vulnerabilities. Ivanti had its monthly patch day, so for May, and it fixed a number of vulnerabilities. One I want to point out here is, and that's a vulnerability that's already exported, exploit it in the wild. However, in order to exploit it, an attacker must first have admin credentials. And as Ivanti points out here, that they patched some vulnerabilities in their January update that would gain access to those credentials. And back then they rightfully recommended that you swap and rotate your credentials. So if you haven't done that yet, well, this puts you at more risk of more persistent exploitation of your devices. So Two things you should do here. Number one, apply the patches. And then of course double check that you recently rotated your credentials on your Ivanti Endpoint Manager mobile devices. That's what's being affected here. Ivanti also published a blog post that sort of went with this security bulletin and stated that they're now well and everybody does really use more AI tools in order to find new vulnerabilities. They are having some initial success and basically that just highlighting that yes, in the next couple months you may see more vulnerabilities being patched. But in some ways it's actually a good thing that Ivanti is using these tools and not just waiting for some researchers kind of to report the vulnerabilities to them, but instead proactively is going out and trying to find as many vulnerabilities as they can in house. I talked earlier this week about Google no longer really disclosing all the vulnerabilities. Sounds like Ivanti is going to disclose all vulnerabilities they find, even the ones they find internally using these AI tools. Well, in users of Redis be aware there is a patch for you and this patch also fixes remote code execution vulnerability. So definitely something that you want to pay attention to. You must have credentials in order to exploit it. So it's not a pre authigation vulnerability. Still in particular with Redis I often see it exposed and so credentials being given to various users. So something that you should pay attention to as a result. They also have a couple of other recommendations which I think are really important for many of these NoSQL databases, like not allowing direct network access, find the strongest authentication method that you can use for the particular database. So also read kind of these footnotes to the advisory in addition to patching the vulnerability. Well, and it's Friday again and we do have another sans edu student to talk about a research project. Marcio, could you introduce yourself please?

[B]

I'm Marcio Enriquez and currently a lead service manager over exposure management in one of the Fortune 500 companies and been doing cybersecurity for more than a decade now in it even longer. So very excited to be here.

[A]

So lots of changes over a decade and you sort of covered a cutting edge kind of topic, you know, AI and hacks around it. So can you explain a little bit what the paper was about?

[B]

So I wanted to focus on something that was relevant and was important to what we're experiencing and seeing every day. And so we were having a meeting actually it was pretty interesting. We had a phishing Alert come in from one of our SOC analyst teams and they're reviewing it and going through. But it quickly escalated when they noticed that the user had clicked on the phishing email, were starting to see their identity across all the machines within the enterprise that ultimately led up to the manager level. We got on a very important call with everybody, including VPs and directors, and we discovered that what occurred was Microsoft's Automatic Attack Disruption tool. It's an autonomous AI tool that's enabled, took these actions without none of us knowing. We didn't even know it was enabled, we didn't even know it was running. And we started to dig deeper and we noticed that the fidelity was actually pretty good because immediately the question from our VP and directors and us were, should we turn it off? Right? Because we had taken actions that are really disabling a user account. And it added the user across the enterprise under the Deny interactive logon. And it did all this like in a short span time, which was really great. And One of our VPs had a question, he said, well, could this be used against us? And that question really provoked my thought process and thus the research topic. I really wanted to dig into that question. You know, from a defense perspective, we're leveraging it or AI almost everywhere, right? Artificial intelligence is like the key word to use in all your defense mechanism tools, especially autonomously, to take automated action increases like your meantime, to respond meantime to closure, you name it, it increases the speed because everything is functioning at machine speed. And so my research delve around, let's look at these defensive tools and see could they be manipulated in a way that would cause a disruption. And thus the topic or the title of my paper, right, is introducing the Autonomous Defense Induced Disruption adid kind of like a mouthful. And how AI driven automated response can be manipulated to disrupt enterprise operations. And so the experiment focused on that really. I worked hard on building out a really good lab, as close to real as possible, E5 trial license, real domain connected to it, a bunch of identities, ran a ton of scripts to create Personas like a user who mimic a user who's barely logging onto the machine, mimic a user that does normal logins, web traffic and whatnot. And really made sure that the E5 license had a lot of telemetry to kind of consume web analysis. Traffic analysis created virtual machines and then created an ATTCK network. So just a KALI Linux box completely isolated and segmented. And the whole purpose was not to check fidelity of the, you know, autonomous actions that were taken. But more importantly, can we Hit the triggers that caused the automated actions to take. Can we simulate that to make it force user containment and how far would it go?

[A]

Yeah, so that reminds me, all of the good old account lockout after five logins.

[B]

Exactly. Password spray is locking everybody out.

[A]

I think Cisco had that feature where it would detect port scans internal to internal.

[B]

Oh yeah.

[A]

And then automatically block systems. And one of the problems there was if you received an email with like lots of image tags and point back to an internal IP address with like colon 80, colon 81, colon 82 and so on.

[B]

Yeah, you got it.

[A]

Trigger that.

[B]

So oops.

[A]

So just by opening the email, you know, you would trigger that because now your email client goes out to all these ports and that internal IP address and. Yeah, and triggers it.

[B]

So in the research again, you know, in the lab environment wasn't focused on like let's fully compromise the system and how to do that. It was more once an attacker gets a foothold into like one machine and you know, through open source intelligence they get a handle of usernames or targeting a few users, ultimately attack disruption. We had about 18 test accounts disabled all 18 accounts. And once I enabled the password write back tool for the hybrid environment, it also locked out my domain admin account on the local environment, which was a lot of fun to get back into. And this showed us one of the big points. The current form of artificial intelligence that we leverage today, and this is IBM, did a really good job of kind of breaking this down, was a artificial narrow intelligence, you know, meaning that the, all the language learning models that we're using, chat, GPTs, clauds, all that every AI that's currently existing is really good at what it's specifically coded or designed to do, meaning contextually it may not be fully comprehend the ramifications of its actions. Case in point, a normal analyst seeing that the domain admin account is doing something malicious, may not so quickly disable that account without first making sure that we have access back into our network so that contextual knowledge appears to be missing. And that was one of the biggest points. Right, Is a call out to that, a call to action, so to speak. Right, let's, we have these tools that we're using, but let's make sure that we have the proper guardrails, break glass accounts, all the traditional things that we do for any type of runbooks that have automatic actions just to ensure we don't get locked out.

[A]

Yeah. Is there any kind of configurability in the tool where you can say okay, don't Lock out these accounts or.

[B]

It was interesting in the beginning of the research, we noticed that there was an option by Microsoft where you can input user identities that you did not want any actions taken on automatically. But an interesting call out is that it's not a requirement to fill in a user to have this enabled. So, you know, if you don't know to look for it, you would never know that it existed. And as my research increased later down the line, I noticed that they added a new feature which was interesting for machines, add machines here, that you don't want containment actions to be taken care of like your crown jewels and things like that. So it's interesting that they're starting to exist, which is great. But if you don't know to configure them, they're not a requirement for having any automatic actions taken.

[A]

And there's no easy undo button or anything like that.

[B]

Actually what I had to do in order to get back into the infrastructure, right into my own lab environment was I had, luckily I had a controlled cloud account in Azure. I had to go in there, log in through that and leverage a live session to undo and disable all the actions that were taken by the autonomous agent. So the rollback can occur. But if I'm targeting anybody just for disruption, I might just target the main IT admin you have.

[A]

What about sort of visibility into what happened? Is there some decent logging that will tell you all?

[B]

The logging is extremely well documented. Everything within the XDR is really, really great. You'll be able to see all the actions that were taken specifically by the autonomous agent and you can roll those back if you see them. And so the bigger point of the paper that I wanted to bring out was that you can see everything happening, right? Everything you can control. Even like if you, let's say you want to roll a few actions back, those are possible. But the bigger danger is that if an adversary were to know what, let's say modern equipment you're using, imagine Apollo Alto or a Fortinet with some sort of autonomous blocking on the edge, right? And they know that how the system works, documentation is great on how the AI works out there. They no longer are trying to hide themselves to cause disruption. They're more like, let's purposefully find these high yield triggers to make actions taken to force a containment that maybe a human would not do. That would cost high severity disruption.

[A]

Yeah, thanks for joining me here. And there should be a link then in the show notes if anybody's interested in the full paper. And yeah, thanks everyone. For listening and talk to you again on Monday. Bye.