Latest / Dark Perimeter: Real Breaches, Real Stakes / Dark Perimeter: "The Forty-Eight Hour Window"
Transcript
- 0:00SPEAKER_00: 88%.
- 0:02SPEAKER_00: That is the number I have not been able to put down all week.
- 0:06SPEAKER_00: In the first half of this year, according to CrowdStrike's 2026 threat hunting report published on August 3rd, 88% of intrusions that followed a publicly disclosed vulnerability happened within 48 hours of that disclosure.
- 0:23SPEAKER_00: Not within the patch cycle, not within the maintenance window, within two days.
- 0:30SPEAKER_00: I'm Cole Draden, this is Dark Perimeter, and tonight we are going to talk about what happens to a security program when the interval between a fix exists and you are already compromised shrinks below the time it takes to schedule a change.
- 0:46SPEAKER_00: With me is Dr.
- 0:47SPEAKER_00: Elliot Vance, who spends his life inside vulnerability data.
- 0:52SPEAKER_00: And Marcus Hale, who has to actually run the change board on Monday morning.
- 0:58SPEAKER_00: Elliot, start with the number itself.
- 1:01SPEAKER_00: What is it measuring and what is it not measuring?
- 1:05SPEAKER_02: It is measuring dwell between disclosure and observed intrusion.
- 1:09SPEAKER_02: And the distinction matters more than the headline does.
- 1:12SPEAKER_02: This is not a claim that 88% of vulnerabilities get exploited.
- 1:16SPEAKER_02: Most never do.
- 1:16SPEAKER_02: It is a claim about the subset that does get exploited.
- 1:20SPEAKER_02: And it says that when exploitation happens at all, it now happens almost immediately.
- 1:25SPEAKER_02: The tail has been cut off.
- 1:26SPEAKER_02: There used to be a long, comfortable slope where a vulnerability would sit for three weeks, then six weeks, then get folded into a commodity exploit kit and turn up an opportunistic scanning.
- 1:37SPEAKER_02: That slope has largely collapsed into a spike at the front.
- 1:40SPEAKER_02: The report also pairs that with a specific case.
- 1:43SPEAKER_02: Two China Nexus groups, which CrowdStrike tracks as VaultPanda and Genesis Panda exploited React 2 Shell within 24 hours of public disclosure.
- 1:53SPEAKER_02: React 2 Shell being the unauthenticated remote code execution issue in React server components and Next.js applications that was disclosed back in December.
- 2:03SPEAKER_02: And separately, the report notes a 42% year-over-year increase in zero-day exploitation from 2024 to 2025.
- 2:12SPEAKER_02: I want to flag one thing about attribution before we build on it.
- 2:15SPEAKER_02: The China Nexus label is CrowdStrike's assessment using their own naming and their own telemetry.
- 2:21SPEAKER_02: The public report does not publish the underlying evidence that supports the nation-state association.
- 2:27SPEAKER_02: That is normal for vendor threat intelligence, and it is not a reason to dismiss it.
- 2:31SPEAKER_02: It is a reason to hold it as a vendor assessment rather than as an adjudicated fact.
- 2:36SPEAKER_01: And I want to be careful about what we do with the number because there is a version of this conversation that is useless.
- 2:42SPEAKER_01: The useless version is patch faster.
- 2:44SPEAKER_01: I have sat in the meeting where an executive reads a statistic like 88%, turns to the infrastructure lead, and says, our SLA is now 24 hours, and then nothing happens because the SLA was never the constraint.
- 2:56SPEAKER_01: The constraint is that you have 400 systems, a change advisory board that meets Tuesdays, an ERP vendor whose support contract voids if you patch out a band, and a plant floor that cannot reboot during a production run.
- 3:07SPEAKER_01: If exploitation happens inside 48 hours, and your realistic organizational patch velocity for a non-trivial system is seven to fourteen days, then the honest conclusion is not go faster.
- 3:18SPEAKER_01: It is that patching has stopped being your primary control for this class of risk, and you need to know what is actually carrying the load in those seven to fourteen days.
- 3:26SPEAKER_00: That is the tension I want to sit in.
- 3:29SPEAKER_00: Marcus, before we get to what carries the load, let's ground this in something that happened this week, rather than in a statistic.
- 3:38SPEAKER_00: Because there is a live example.
- 3:40SPEAKER_00: Enable NCentral.
- 3:43SPEAKER_00: Elliot, walk us through it.
- 3:54SPEAKER_02: Agent on the endpoint, high privilege, remote control, script execution.
- 3:58SPEAKER_02: It is by design the thing that has hands on everything.
- 4:01SPEAKER_02: The sequence is this there was an authentication bypass, CVE 2026, Waiting 556, CVSS 8.2.
- 4:10SPEAKER_02: It got patched.
- 4:11SPEAKER_02: Then it turned out the patch was incomplete, and the incomplete patch produced a second vulnerability: CVE 2026 Waite 577, also 8.2, also an authentication bypass by an alternate path.
- 4:24SPEAKER_02: On August 2nd, Enable acknowledged that a limited number of customers had been compromised and the fixed shipped in version 2026-3, HF1.
- 4:32SPEAKER_02: On August 3rd, SISA added CVE 2026, Waiteen 577, to the known exploited vulnerabilities catalog, with a federal remediation deadline of today, August 6th.
- 4:44SPEAKER_02: Huntress reported observing exploitation across multiple organizations, and their characterization was that they had not seen a broad indiscriminate campaign.
- 4:53SPEAKER_02: So this reads as targeted rather than sprayed, at least as of that reporting.
- 4:57SPEAKER_02: The post-exploitation behavior described as reconnaissance, process enumeration, lateral movement toward domain controllers, and persistence established by abusing the default MSP support username used for take control sessions.
- 5:24SPEAKER_01: That is the attacker using the tool the way the tool is documented to work, and it means your detection has to distinguish between the vendor's support tooling doing support and an intruder doing the same thing with the same account from the same console.
- 5:36SPEAKER_01: Most shops cannot make that distinction because they never baselined it.
- 5:39SPEAKER_01: And the incomplete patch part is the piece I would put in front of leadership.
- 5:43SPEAKER_01: Your organization did the right thing.
- 5:44SPEAKER_01: You applied 18556 inside your window, you closed the ticket, you reported green on the dashboard, and you were still exposed because the fix did not fully fix it.
- 5:53SPEAKER_01: There is no amount of patch discipline that protects you from a patch that does not work.
- 5:57SPEAKER_01: That risk does not live on your side of the boundary at all.
- 6:01SPEAKER_00: Which raises the vendor concentration question.
- 6:04SPEAKER_00: If the RMM is the thing with hands on everything, then a bypass in the RMM is not one company's incident.
- 6:13SPEAKER_01: It is every downstream client of every MSP that runs it, and most of those clients have no idea the product exists.
- 6:19SPEAKER_01: If you are a 40-person law firm and you have outsourced IT, your actual attack surface includes your provider's end central instance, your provider's credential hygiene, and your provider's patch velocity, none of which appear anywhere in your risk register.
- 6:32SPEAKER_01: This is the third-party risk problem in its purest form, and the honest position is that a vendor questionnaire does not touch it.
- 6:39SPEAKER_01: What touches it is a contractual right to be notified, a defined notification window, and knowing which of your vendors hold standing privileged access into your environment.
- 6:48SPEAKER_01: That list is usually shorter than people expect, and almost nobody maintains it.
- 6:53SPEAKER_02: It is also worth naming the rest of the week's catalog editions because the pattern holds.
- 6:58SPEAKER_02: Alongside the two N central entries, SISA added CVE 2026-63077, a deserialization of untrusted data issue in JetBrain's Team City, CVE 2026-34486 in Apache Tomcat, categorized as missing encryption of sensitive data, and CVE 2026-9198, a code injection flaw in IBM Langflow.
- 7:22SPEAKER_02: Look at that list as a set: a remote management platform, a build server, an application server, and an AI application framework.
- 7:30SPEAKER_02: Three of those four sit inside the software delivery path or the operations path rather than at the edge.
- 7:36SPEAKER_02: The perimeter framing keeps failing because the valuable positions are not at the perimeter.
- 7:42SPEAKER_00: Which brings us to the third thing that happened this week, and I think it is the most structurally interesting one.
- 7:48SPEAKER_00: Elliott.
- 7:50SPEAKER_00: Chaindrop.
- 7:51SPEAKER_02: On August 4th, Elastic Security Labs published analysis of a self-propagating worm in the NPM ecosystem that they call Chaindrop, and the reporting is that it compromised more than 400 packages.
- 8:04SPEAKER_02: The entry point per Elastic was the compromise of the maintainer of Kiev.
- 8:08SPEAKER_02: The propagation is the part worth understanding.
- 8:10SPEAKER_02: The malware activates through pre-install hooks.
- 8:13SPEAKER_02: Once running, it enumerates every package the victim has published rights to, injects itself into each one, and republishes them to the registry.
- 8:21SPEAKER_02: Elastic notes that propagation requires an NPM token that carries both package write permission and a two-factor bypass capability.
- 8:30SPEAKER_02: So it spreads along the graph of publishing authority, which is a completely different shape from spreading along a network.
- 8:36SPEAKER_02: The collection component scans for something north of 300 credential patterns, cloud provider keys across AWS, GCP, Azure, and Alibaba, GitHub tokens, SSH keys, Kubernetes service accounts, NPM tokens, Hachi Corp Vault tokens, and notably a whole category of AI tooling credentials, Anthropic, OpenAI, Gemini Cursor, and others.
- 9:01SPEAKER_02: Stolen material is encrypted with AES256 GCM under an RSA wrapped key.
- 9:07SPEAKER_02: For command and control, it queries an Ethereum smart contract at runtime to retrieve the current exfiltration endpoint, with fallbacks that include searching GitHub commit history and exfiltrating through repositories under the victim's own account.
- 9:21SPEAKER_02: Elastic assesses this as the same actor lineage as the earlier Chai Hulude campaigns, based on Dune themed strings and similarities in the payload.
- 9:31SPEAKER_02: That is a code similarity assessment, which is inferential.
- 9:34SPEAKER_02: I would report it as Elastic's assessment rather than as established identity.
- 9:39SPEAKER_00: The Ethereum piece is the one I keep turning over.
- 9:42SPEAKER_00: Explain why that matters operationally rather than as a novelty.
- 9:47SPEAKER_02: Because it removes the takedown.
- 9:49SPEAKER_02: Traditional C2 resilience is domain generation algorithms, fast flux, bulletproof hosting.
- 9:55SPEAKER_02: All of those still have a point where a registrar or a host can be leaned on.
- 9:59SPEAKER_02: A read from a public blockchain has no registrar.
- 10:02SPEAKER_02: There is nobody to serve.
- 10:03SPEAKER_02: The contract is a permanent globally readable configuration channel that the operator can update at will and that a defender can observe but cannot revoke.
- 10:12SPEAKER_02: You can block the RPC endpoints your environment uses to reach it, and you should, but you cannot remove the resource.
- 10:19SPEAKER_01: And the credential list is the thing I want practitioners to hear because it tells you what the actual loss event is.
- 10:25SPEAKER_01: Nobody is stealing your NPM packages for their own sake.
- 10:28SPEAKER_01: They are stealing the keys that were sitting in your CI runner's environment, which means the blast radius of a compromised dependency is not our build broke.
- 10:36SPEAKER_01: It is every secret that any build job has ever had in scope.
- 10:39SPEAKER_01: If your CI service account has a long-lived cloud key with broad permissions, and it does, because it always does, then a package compromise three layers deep in your dependency tree is a cloud account compromise.
- 10:49SPEAKER_01: The AI tooling credentials being a name priority target is new, and I do not think most programs have caught up.
- 10:55SPEAKER_01: Developers now hold API keys to model providers.
- 10:58SPEAKER_01: Those keys frequently sit in plain text config in the home directory.
- 11:01SPEAKER_01: They are usually not in the enterprise secret store.
- 11:04SPEAKER_01: They are usually not rotated, and in a lot of organizations they are not even inventoried because they were expensed on a card rather than procured.
- 11:11SPEAKER_01: That is an entire credential class that exists outside the identity program.
- 11:15SPEAKER_00: So let's do the practical turn.
- 11:18SPEAKER_00: Given a forty-eight-hour exploitation window, an incomplete vendor patch, and a worm that travels through publishing rights rather than through the network, what does a program actually change on Monday?
- 11:32SPEAKER_00: Marcus, you first, and be specific.
- 11:35SPEAKER_01: Four things in the order I would do them.
- 11:37SPEAKER_01: First, build the list of vendors and products that hold standing privileged access into your environment.
- 11:42SPEAKER_01: RMM, backup agents, endpoint management, monitoring agents, anything with a service running a system everywhere.
- 11:48SPEAKER_01: That list is your real crown jewel inventory, and it is not your data classification exercise.
- 11:53SPEAKER_01: For each one, know your notification path and know how you would revoke that access in an hour if you had to.
- 11:58SPEAKER_01: Most people have never tested revocation.
- 12:00SPEAKER_01: Second, stop measuring patch compliance and start measuring exposure time on the subset that matters.
- 12:06SPEAKER_01: Compliance percentage across 10,000 endpoints is a number that makes leadership comfortable and tells you nothing.
- 12:11SPEAKER_01: Time to remediate on internet facing and unprivileged access systems, tracked against CISA Kev specifically, is a number that maps to the risk we have been describing.
- 12:20SPEAKER_01: Kev is a small list.
- 12:21SPEAKER_01: It is achievable to be current against it.
- 12:23SPEAKER_01: Being current against Kev is worth more than being at 94% against everything.
- 12:27SPEAKER_01: Third, get the compensating controls named explicitly, because they are what covers the 7 to 14 days you cannot compress.
- 12:33SPEAKER_01: Network segmentation around management planes, so the RMM console is not reachable from a user VLAN.
- 12:39SPEAKER_01: Egress control so a build runner cannot reach an arbitrary endpoint.
- 12:43SPEAKER_01: Behavioral detection on your privileged service accounts, which means baselining what the vendor's support account normally does so that abnormal use of it is visible.
- 12:51SPEAKER_01: Say these out loud in the risk acceptance so that when a patch is late, everyone knows what is carrying the weight instead of pretending nothing is.
- 12:58SPEAKER_01: Fourth, and this is the CI1, kill long-lived credentials in build pipelines.
- 13:03SPEAKER_01: Short-lived workload identity federated tokens scope to the job.
- 13:06SPEAKER_01: If Chaindrop had landed in a pipeline whose credentials expire in 15 minutes and are scoped to one repository, the loss event is small.
- 13:14SPEAKER_01: If it lands in a pipeline holding a static admin key, the loss event is the environment.
- 13:19SPEAKER_02: I would add one that is analytical rather than operational.
- 13:22SPEAKER_02: Change what you consider a patch state.
- 13:24SPEAKER_02: The NCentral case is a demonstration that the vendor issued a fix and we applied it is not a terminal state.
- 13:30SPEAKER_02: Incomplete patches are common and under tracked.
- 13:33SPEAKER_02: Practically that means when a critical authentication or remote code execution issue is fixed in a product that sits in your privileged tier, keep monitoring that CVE's family for 30 to 90 days rather than closing the file.
- 13:46SPEAKER_02: A second CVE that supersedes the first is a signal that the original analysis was wrong, and it should reopen your assessment rather than start a new one from zero.
- 13:58SPEAKER_00: Because I think they are one story rather than three.
- 14:08SPEAKER_00: The NCentral case tells us that even correct human response inside that interval can fail because the fix itself can be incomplete, and that the most valuable position for an attacker is the tool you installed to manage everything.
- 14:24SPEAKER_00: ChainDrop tells us that propagation has stopped following the network topology we defend and started following the trust topology we never mapped, publishing rights, tokens, dependency graphs.
- 14:38SPEAKER_00: The common thread is that all three of these bypass the assumption underneath most security programs, which is that there is a window in which a person notices and decides.
- 14:49SPEAKER_00: Elliot, is that the right read?
- 14:53SPEAKER_02: I think it is, with one qualification.
- 14:55SPEAKER_02: It is not that human decision making becomes irrelevant.
- 14:58SPEAKER_02: It is that it has to move upstream.
- 15:00SPEAKER_02: You cannot decide during the incident anymore, so you decide beforehand, and what you are deciding is architecture.
- 15:06SPEAKER_02: Where the trust boundaries sit, which credentials can exist in which places, what is reachable from what.
- 15:12SPEAKER_02: Those decisions execute at machine speed because they are already encoded when the event happens.
- 15:16SPEAKER_02: The organizations that are handling this well are not the ones with the fastest patch cycles.
- 15:21SPEAKER_02: They are the ones whose architecture makes a late patch survivable.
- 15:25SPEAKER_01: And the thing I would say to anyone running a real program is that none of this is a reason to panic, and it is definitely not a reason to buy something.
- 15:32SPEAKER_01: Every control I listed is something you already own or can configure, segmentation, egress filtering, credential lifetime, an accurate inventory of who has privileged access into your environment.
- 15:42SPEAKER_01: It is unglamorous, and it is mostly project work rather than product work, which is exactly why it does not get done.
- 15:48SPEAKER_01: But the 48-hour window is not a problem you can purchase your way out of.
- 15:51SPEAKER_01: It is a problem you design your way out of slowly in advance.
- 15:55SPEAKER_00: The window closed while we were still writing the change request.
- 16:00SPEAKER_00: That is the honest summary of where we are in August of 2026.
- 16:05SPEAKER_00: The fix is not to write the change request faster.
- 16:09SPEAKER_00: It is to build something that does not depend on the change request landing in time.
- 16:15SPEAKER_00: Elliot, Marcus, thank you both.
- 16:18SPEAKER_00: For everything we covered tonight, the primary sources are the CrowdStrike 2026 Threat Hunting Report from August 3rd, and ABLE's own advisory, and the CESA known exploited vulnerabilities catalog entries from August 3rd.
- 16:34SPEAKER_00: And the Elastic Security Labs analysis of ChainDrop from August 4th.
- 16:40SPEAKER_00: Where we described attribution, we attributed it to the organization making the assessment because that is what it is.
- 16:48SPEAKER_00: This is dark perimeter.