Latest / Dark Perimeter: Real Breaches, Real Stakes / Your Tier Zero List Is Wrong
Transcript
- 0:00SPEAKER_02: Three things happened in the last week that look like three different stories.
- 0:04SPEAKER_02: A network appliance vendor shipped an emergency bulletin for bugs that were already being exploited.
- 0:10SPEAKER_02: A source code platform patched a 9.9 in the piece of software that runs its AI agents.
- 0:17SPEAKER_02: And a storage vendor pushed six critical fixes for the component that connects enterprise arrays to Kubernetes.
- 0:25SPEAKER_02: Different vendors, different products, different weeks of somebody's life.
- 0:30SPEAKER_02: I want to argue they are one story, and that the story is about a category of asset most security programs cannot produce a list of on request.
- 0:41SPEAKER_02: I am Cole Draden.
- 0:42SPEAKER_02: With me are Dr.
- 0:43SPEAKER_02: Elliot Vance and Marcus Hale.
- 0:46SPEAKER_02: Elliot start with the one that is actually on fire.
- 0:50SPEAKER_00: That is Citrix Netscaler, and the sequence matters more than the bugs.
- 0:54SPEAKER_00: There are two primary vulnerabilities.
- 0:56SPEAKER_00: CVE 2026-88771 is an unauthenticated command injection.
- 1:02SPEAKER_00: The appliance fails to validate input, and an actor with no credentials can run commands against it.
- 1:08SPEAKER_00: CVE 2026-88772 is a memory overflow affecting DTLS configurations that can produce either remote code execution or denial of service.
- 1:19SPEAKER_00: Both carry a C VSS version 4 base score of 9.5.
- 1:23SPEAKER_00: Citrix published its bulletin on the 29th of September, and SISA added both to the known exploited vulnerabilities catalog the same day.
- 1:31SPEAKER_01: And the exploitation predates all of that, which is the part that should ruin somebody's quarter.
- 1:36SPEAKER_01: Grey noise reported attempts on the 26th, and in their own words, there were no detections available specific to the CVE at that point.
- 1:43SPEAKER_01: Citrix started telling customers to disconnect servers on the twenty eighth.
- 1:47SPEAKER_01: So the order of events was attacked, told to unplug, then told why.
- 1:51SPEAKER_01: If your detection strategy is to wait for a signature, you had a three-day window where the correct answer was unavailable to you by construction.
- 1:58SPEAKER_02: Marcus.
- 2:00SPEAKER_02: How many organizations are we talking about?
- 2:03SPEAKER_02: I have seen two numbers this week, and they do not agree.
- 2:08SPEAKER_01: They do not, and I would rather say that out loud than pick the scarier one.
- 2:11SPEAKER_01: Palo Alto's unit 42 put potentially vulnerable exposed instances at 50,277 as of the twenty-seventh.
- 2:19SPEAKER_01: The Shadow Server Foundation counted more than twenty thousand visible and potentially vulnerable.
- 2:24SPEAKER_01: Those are different methodologies measuring different things, probably fingerprinting versus exposure, and I do not know which is closer to a defensible number.
- 2:32SPEAKER_01: What I will say is that both are large and neither is the number that matters to you.
- 2:35SPEAKER_01: The number that matters to you is one, and it is whether yours is in there.
- 2:39SPEAKER_00: On impact rather than exposure, Mandiant's Charles Carmicall said dozens of organizations have been affected across North America and Europe, spanning government, financial services, education, telecommunications, and legal.
- 2:53SPEAKER_00: That is reporting from an incident response firm with visibility into actual engagements, not a scan.
- 2:59SPEAKER_00: It is a different kind of claim than a port count, and it should be weighted differently in both directions.
- 3:04SPEAKER_00: It is better evidence of real harm, and it is a floor, not a total.
- 3:09SPEAKER_02: Elliot, what did the attackers actually do once they were in?
- 3:13SPEAKER_02: Because I think that is where the operational lesson lives.
- 3:16SPEAKER_00: Unit 42 documented two distinct web shell strategies.
- 3:20SPEAKER_00: One delivered a Debian package through the DTLS path, dropping PHP shells with command and control encrypted using RC4 keyed on an MD5 hash of a hard-coded passphrase.
- 3:32SPEAKER_00: The other was a three-stage PHP injection chain, a base 64 dropper hidden in user agent headers, log poisoning through deliberately failed login attempts, and then a Perl script that extracted and ran the stage commands.
- 3:45SPEAKER_00: Both ended up with persistence in web accessible directories and modified Apache configuration.
- 3:51SPEAKER_00: The fingerprinting, incidentally, appears to go back to the 21st of August.
- 3:56SPEAKER_01: Which brings us to the sentence every executive needs to hear in the exact words the responder used.
- 4:01SPEAKER_01: KarmaCall's guidance was that customers should check whether they have been compromised before upgrading, because patching without the other steps might not resolve the infection.
- 4:08SPEAKER_01: Unit 42 put it even more plainly.
- 4:10SPEAKER_01: Updating and patching will not remove access for attackers who have already established persistence.
- 4:15SPEAKER_01: I want to sit on that for a second, because there is a reflex in this industry that treats patch deployed as incident closed.
- 4:21SPEAKER_01: On this one, patch deployed is evidence destroyed if you did it before you looked.
- 4:25SPEAKER_02: That is a real tension for a practitioner, though.
- 4:29SPEAKER_02: The patch is the thing you can do in an hour, and the compromise assessment is the thing that takes a week and a vendor contract.
- 4:38SPEAKER_02: What do you actually tell people to do?
- 4:41SPEAKER_01: You do both in an order, and you write the order down before you are in the middle of it.
- 4:45SPEAKER_01: Preserve first, take the configuration, take the logs, take a snapshot of the appliance if your platform lets you, and get it somewhere the appliance cannot reach.
- 4:53SPEAKER_01: Then mitigate, which may mean taking the thing out of the path rather than patching it in place.
- 4:57SPEAKER_01: Then hunt, specifically for sessions that survived a credential change, for files and web accessible directories that nobody can account for, and for a patchy configuration that does not match your build.
- 5:07SPEAKER_01: Then patch, if you only have capacity for two of those four, preserve and mitigate, beat patch and hope.
- 5:13SPEAKER_00: There is also a third Netscaler item from this week that is easy to conflate and should not be.
- 5:18SPEAKER_00: CVE 2026-88779 is a separate memory overflow, and per Citrix the consequence is denial of service on unmitigated deployments rather than code execution.
- 5:29SPEAKER_00: It requires SIML configured as either service provider or identity provider, together with gateway or AAA functionality on premises.
- 5:38SPEAKER_00: CISA added it to the catalog on the 5th of October with a federal remediation deadline of the 7th.
- 5:44SPEAKER_00: Researchers reproduced it, and honeypots observed probing from multiple addresses, with attempts to install web shells reported and malware binaries found on some systems.
- 5:54SPEAKER_00: Denial of service is a materially smaller claim than unauthenticated remote code execution, and I want the distinction preserved.
- 6:05SPEAKER_02: Because this is the one I think is genuinely new.
- 6:09SPEAKER_02: GitLab patched a 9.9 in its AI gateway.
- 6:14SPEAKER_02: Elliot.
- 6:15SPEAKER_02: What is the mechanism?
- 6:16SPEAKER_00: CVE 2026 90970 disclosed on the 2nd of October CVSS 9.9.
- 6:24SPEAKER_00: The flaws in the prompt template of custom flows within the Duo Agent platform.
- 6:29SPEAKER_00: A logged in user could escape the prompt template sandbox through a specially crafted flow configuration, and that escape could lead to arbitrary command execution on the gateway itself.
- 6:39SPEAKER_00: Fixed in gateway versions 19.2.4, 19.3.2, and 19.4.1.
- 6:46SPEAKER_00: It affects organizations running their own AI gateway.
- 6:49SPEAKER_00: GitLab.com, GitLab dedicated, and self-managed instances that use GitLab's hosted gateway are not affected.
- 6:56SPEAKER_00: GitLab notified self-hosted customers before publishing, and as of the advisory, there is no confirmed exploitation.
- 7:04SPEAKER_01: I want to be careful not to oversell this one because the failure mode of a show like this is treating every 9.9 as equivalent.
- 7:10SPEAKER_01: This needs authentication.
- 7:12SPEAKER_01: It needs access to a specific platform feature.
- 7:14SPEAKER_01: Nobody has confirmed it being used.
- 7:16SPEAKER_01: Set against an unauthenticated command injection with web shells already on boxes.
- 7:20SPEAKER_01: It is a lower urgency item this week.
- 7:22SPEAKER_01: And it is still the most interesting thing that happened because of what the gateway is.
- 7:26SPEAKER_02: Say more about that.
- 7:28SPEAKER_01: An AI gateway in a software development platform is not an application.
- 7:33SPEAKER_01: It is a thing that holds credentials to your repositories and your model endpoints and runs on your network so that agents can do work on behalf of your developers.
- 7:40SPEAKER_01: It is functionally a service account with a shell.
- 7:43SPEAKER_01: If somebody gets command execution there, the question is not what data was in the gateway.
- 7:47SPEAKER_01: The question is what the gateway was allowed to reach.
- 7:50SPEAKER_01: And in most deployments, I have seen the honest answer is nearly everything, because restricting it would have made the agent less useful, and the agent was the point.
- 7:58SPEAKER_00: The escape vector is the part I would underline for anyone building in this space.
- 8:02SPEAKER_00: The boundary that failed was a prompt template sandbox.
- 8:06SPEAKER_00: That is a boundary between content and execution, and the content side of it is authored by users as configuration rather than as code.
- 8:14SPEAKER_00: We have 40 years of institutional knowledge about validating input that arrives as data and roughly four years of practice at a system where the input is instructions.
- 8:24SPEAKER_00: A prompt template is a program that does not look like a program, so it does not get a program's review.
- 8:29SPEAKER_00: I expect this specific class, configuration that is secretly executable, to produce a long series of these.
- 9:03SPEAKER_01: There is a remote privilege escalation to root on cluster nodes.
- 9:06SPEAKER_01: There is a compromise of administrative access to the authorization proxy itself.
- 9:10SPEAKER_01: There is authentication token forgery for administrator rights, and there is a Kubernetes access control bypass permitting clusterwide reading of secrets, fixed in 1.18.0.
- 9:20SPEAKER_01: Dell had not identified these as actively exploited at publication.
- 9:25SPEAKER_00: The affected integration covers PowerStore, PowerScale, PowerFlex, PowerMax, and Unity XT connected to Kubernetes.
- 9:33SPEAKER_00: The combination that matters is the last two items on that list.
- 9:36SPEAKER_00: Cluster wide read of secrets plus token forgery is not a storage bug.
- 9:41SPEAKER_00: Secrets are where the cluster keeps the credentials for everything else it talks to.
- 9:45SPEAKER_00: So reading all of them is a credential harvest with a storage CVE number attached to it.
- 9:51SPEAKER_02: So here is the thesis I want to put to both of you.
- 9:54SPEAKER_02: Three stories.
- 9:56SPEAKER_02: Three vendors, and in every case, the compromised thing was not where the data lives.
- 10:01SPEAKER_02: It was the thing that administers where the data lives.
- 10:04SPEAKER_02: Is that a real pattern, or am I drawing a line through three points?
- 10:09SPEAKER_01: It is real, and I can tell you exactly how it happens because I have watched it happen.
- 10:13SPEAKER_01: Go ask your team for the tier zero asset list.
- 10:15SPEAKER_01: You will get domain controllers.
- 10:17SPEAKER_01: If the program is mature, you will get the identity provider and the privilege access management system.
- 10:22SPEAKER_01: You will almost never get the VPN appliance, even though it terminates authentication for the entire remote workforce.
- 10:28SPEAKER_01: You will not get the CICD system, even though it can deploy code to production without a human.
- 10:32SPEAKER_01: You will not get the Kubernetes storage driver, even though it can read every secret in the cluster.
- 10:37SPEAKER_01: Those things were installed by infrastructure, networking, and platform teams as plumbing, and plumbing does not get classified as an identity system.
- 10:44SPEAKER_00: The structural reason is that our classification scheme is about data sensitivity, and the risk is about reachability.
- 10:51SPEAKER_00: We built tiering models that ask what is stored here.
- 10:54SPEAKER_00: The question that predicts this week's incidence is what can this reach, and with whose authority?
- 10:59SPEAKER_00: An appliance that stores nothing and holds the SAML keys for everything scores low on the first question and highest possible on the second.
- 11:07SPEAKER_00: Until the inventory asks the second question, these assets will keep being surprises.
- 11:12SPEAKER_02: Marcus, make that concrete.
- 11:15SPEAKER_02: Somebody is listening to this on the way into work and they run a program.
- 11:20SPEAKER_02: What do they do this week?
- 11:22SPEAKER_01: Four things.
- 11:23SPEAKER_01: And the first one is not patching.
- 11:25SPEAKER_01: One, produce the list.
- 11:26SPEAKER_01: Every system that holds credentials for or can execute against another system.
- 11:30SPEAKER_01: Appliances, gateways, CICD, secret stores, storage drivers, backup servers, hypervisor managers, anything with an agent on every host.
- 11:38SPEAKER_01: If you cannot produce it in a day, that is your actual finding, and it is more valuable than any individual CVE on this episode.
- 11:44SPEAKER_01: 2.
- 11:44SPEAKER_01: For everything on that list, write down what it can reach and under what identity, and then cut what is not needed.
- 11:50SPEAKER_01: Most of these things have standing access to far more than their function requires because scoping was deferred at install and never revisited.
- 11:56SPEAKER_01: 3.
- 11:57SPEAKER_01: Decide now, in writing, what your preserved and patch order is, so that the next time a vendor says disconnect this immediately, you are not inventing evidence handling under pressure.
- 12:05SPEAKER_01: 4.
- 12:05SPEAKER_01: Patch on the vendor's timeline with the Kev dates as your floor rather than your target.
- 12:10SPEAKER_00: I would add one measurement to that.
- 12:12SPEAKER_00: For each asset on the list, record how you would know if it were compromised, not whether you have a tool, how you would know.
- 12:18SPEAKER_00: Several of this week's items are appliances where you do not get a shell, your endpoint agent does not run, and the only telemetry is what the vendor chose to emit.
- 12:27SPEAKER_00: If the honest answer for an asset is that you would find out when somebody else tells you, that is a decision you are making rather than a gap you have.
- 12:35SPEAKER_00: And it should be a decision somebody senior signs.
- 12:48SPEAKER_02: Elliot, what is established and what is not?
- 12:52SPEAKER_00: What is reported is that a suspect connected to the Shiny Hunters Extortion Group was detained in Jordan on the 29th of September, known by the Handel Ray, and that this person is reportedly cooperating with the FBI to identify other members associated with a series of data breaches and extortion campaigns.
- 13:10SPEAKER_00: I would treat the cooperation specifically as reporting rather than as confirmed, and nothing here has been adjudicated.
- 13:17SPEAKER_00: One detention does not dismantle a group that operates as a loose brand.
- 13:21SPEAKER_00: And the historical pattern after this kind of event is reconstitution under a different name rather than cessation.
- 13:28SPEAKER_01: It matters anyway, and not for deterrence.
- 13:30SPEAKER_01: It matters because cooperation produces attribution, attribution produces indicators, and indicators show up in your feeds three to six months later as the reason you can finally name something you have been seeing in your logs for a year.
- 13:42SPEAKER_01: The enforcement outcome and the defensive outcome run on completely different clocks, and the defensive one is the one that eventually reaches your console.
- 13:49SPEAKER_02: Let me close where we started.
- 13:52SPEAKER_02: The three vulnerabilities on this episode have almost nothing in common technically.
- 13:58SPEAKER_02: A command injection in an appliance, a sandbox escape in a prompt template, an authorization bypass in a storage driver.
- 14:07SPEAKER_02: What they share is a position in the architecture.
- 14:11SPEAKER_02: Each one sits above the things we spend our budget protecting and holds the keys to them, and in each case the thing was installed by somebody solving a different problem than the one we are now having.
- 14:24SPEAKER_02: The patches for all of this will be deployed in the next two weeks by competent people.
- 14:29SPEAKER_02: The list will not get made because nobody is going to be fired for not having it.
- 14:35SPEAKER_02: That is the gap, and it is an inventory problem wearing three C V E numbers this week.
- 14:42SPEAKER_02: This is dark perimeter.