Latest / Dark Perimeter: Real Breaches, Real Stakes / Dark Perimeter: "The Machine Picked the Target"
Transcript
- 0:00SPEAKER_00: Six hundred and forty seven thousand and seventeen.
- 0:04SPEAKER_00: That is the number of internet exposed N eight N workflow automation instances that an attacker's software counted on its own before deciding whether any of them were worth attacking.
- 0:17SPEAKER_00: It narrowed that to twenty-five thousand two hundred and nine in one country, sampled about a hundred of those, read the CVE, checked how popular the project was on GitHub, and then wrote in its own session log, and I'm quoting a machine here.
- 0:34SPEAKER_00: The N8N1 with 258 stars and CVSS 10.0 looks extremely promising.
- 0:42SPEAKER_00: That log is public now because the agent running the operation started a web server in its own home directory and served its operator's entire toolkit to the internet.
- 0:53SPEAKER_00: Palo Alto's Unit 42 walked in the open front door and took notes.
- 0:59SPEAKER_00: I am Cole Draden.
- 1:01SPEAKER_00: This is Dark Perimeter with me, as always, Dr.
- 1:04SPEAKER_00: Elliot Vance, who is going to tell us what that report actually establishes and what it does not.
- 1:12SPEAKER_00: And Marcus Hale, who is going to tell us what changes on Monday morning.
- 1:17SPEAKER_00: Which I suspect is less than the headlines suggest and more than we would like.
- 1:22SPEAKER_00: Last week we talked about the 48-hour window about response time collapsing.
- 1:29SPEAKER_00: This week is not about speed.
- 1:31SPEAKER_00: This week is about who is choosing the target.
- 1:34SPEAKER_01: Let me set the boundaries of the claim first, because this is a story that gets oversold in both directions.
- 1:40SPEAKER_01: Unit 42 published this on July 30th and updated it on August 11th.
- 1:45SPEAKER_01: The actor operates under the handles Nate and Kan Yuan, is assessed to be based in Zhuhai, China, and describes himself publicly as a binary security researcher.
- 1:54SPEAKER_01: He runs an automated vulnerability intelligence pipeline he calls One Day News, which scrapes newly disclosed remote code execution bugs and pushes alerts out over Telegram.
- 2:04SPEAKER_01: That part is not new.
- 2:05SPEAKER_01: Bug feeds have existed for 20 years.
- 2:08SPEAKER_01: What is new is the layer on top.
- 2:10SPEAKER_01: He wired DeepSeek as the reasoning engine into an orchestration framework called Hermes Agent, which gave the model a terminal, a telegram command and control channel, and a set of custom skills.
- 2:21SPEAKER_01: And then he pointed it at the internet.
- 2:22SPEAKER_01: The agent enumerated targets through FOFA, which is essentially a Chinese showdan.
- 2:27SPEAKER_01: It assessed and prioritized vulnerabilities.
- 2:30SPEAKER_01: It downloaded public exploit code on its own initiative.
- 2:34SPEAKER_01: It attempted exploitation.
- 2:35SPEAKER_01: And when a target class did not pan out, it pivoted to a different CVE based on its own read of deployment scale.
- 2:42SPEAKER_01: That is the full loop, and it ran without a human in it.
- 2:45SPEAKER_00: And the OPSEC failure that gave us all of this?
- 2:48SPEAKER_01: The agent started a file server in its home directory.
- 2:51SPEAKER_01: Tool configurations, API keys, exploit scripts, session logs, all of it exposed.
- 2:58SPEAKER_01: Unit 42 recovered a single complete Hermes session, dated May 7th.
- 3:02SPEAKER_01: That is the artifact that lets us say any of this with confidence rather than inference.
- 3:07SPEAKER_01: I want to be precise.
- 3:09SPEAKER_01: We are not reasoning from telemetry and guessing at intent.
- 3:12SPEAKER_01: We have the operator's own environment.
- 3:15SPEAKER_02: Okay, but let me be the person in the room who has to brief a board on this, because there is a number in that report that is doing a lot of work and almost nobody is quoting it.
- 3:23SPEAKER_02: What did the autonomous part actually accomplish?
- 3:26SPEAKER_01: Nothing.
- 3:27SPEAKER_02: Say that again, slower.
- 3:29SPEAKER_01: The autonomous exploitation attempts failed.
- 3:32SPEAKER_01: The agent went after Langflow, CVE 2026-33011, and did not get in because autologin was disabled on the targets.
- 3:40SPEAKER_01: It went after N8N, CVE 2026-21858, and CVE 225-68613, one of which is a CVSS 10.0, and did not get in because authentication was required.
- 3:54SPEAKER_01: Unit 42 says it plainly, the autonomous attempts did not achieve full compromise of any of their intended targets.
- 4:01SPEAKER_02: And the compromises that did happen?
- 4:04SPEAKER_01: All manual.
- 4:05SPEAKER_01: Every confirmed impact in that report has a human hand on it.
- 4:08SPEAKER_01: Data exfiltration from three Citrix Netscaler targets via CVE 2026-3055.
- 4:15SPEAKER_01: Command execution confirmed on 11 Marimo Notebook instances via CVE 2026-39987.
- 4:23SPEAKER_01: Java deserialization attempts against nine Apache Tomcat servers, that is CVE 2026-34486, which we will come back to.
- 4:33SPEAKER_01: Reverse shell callbacks against three IKE VPN endpoints.
- 4:37SPEAKER_01: And a sustained multi-day operation against a Malaysian government entity with proxy anonymization added on later attempts, which is a very human thing to do after you realize you were noisy the first time.
- 4:49SPEAKER_00: So the headline writes itself as AI launches autonomous cyberattack.
- 4:56SPEAKER_00: And the actual finding is AI launched autonomous cyber attack and was bad at it, while its owner did the real damage by hand.
- 5:06SPEAKER_02: Which is the correct read of the evidence and completely the wrong lesson to walk away with.
- 5:09SPEAKER_02: Cole, I have sat in the meeting where somebody says the AI attacks failed, so we have time.
- 5:14SPEAKER_02: That meeting is how you end up 18 months behind.
- 5:17SPEAKER_02: Read the sentence, Unit 42 actually wrote at the end because they were careful with it.
- 5:21SPEAKER_01: While the observed campaign had limited impacts, the workflow confirms a functional, end-to-end autonomous offensive capability.
- 5:48SPEAKER_02: That part now runs unattended.
- 5:50SPEAKER_02: The exploitation failed because those specific targets were configured correctly.
- 5:54SPEAKER_02: Configuration drift is a fact of the universe.
- 5:56SPEAKER_02: The targeting worked.
- 5:57SPEAKER_00: Elliot.
- 5:59SPEAKER_00: There is a detail in the toolkit inventory I want you to unpack because I think it says something about where this is going.
- 6:08SPEAKER_00: He was not just running DeepSeq.
- 6:10SPEAKER_01: He was evaluating a market.
- 6:12SPEAKER_01: Alongside the Hermes and DeepSeq stack, which did the autonomous work, the environment showed limited usage of Claude Code running Opus through a proxy, Codex running GPT-5.4 through a proxy, and several Chinese market models, Quencode, GLM, Kimi, Minimax.
- 6:30SPEAKER_01: The Western tools were routed through a proxy service to reduce traceability, and anti-attribution settings were configured across all of them.
- 6:37SPEAKER_01: Unit 42 is explicit that they cannot verify actual usage of the Codex path because the operator configured the system to limit local response storage.
- 6:47SPEAKER_01: They flagged the correlation between trusted directories and successful campaigns as notable.
- 6:52SPEAKER_01: Notable is doing honest work there.
- 6:54SPEAKER_01: It is an assessment, not a finding.
- 6:56SPEAKER_00: And the provider side?
- 7:13SPEAKER_01: So the guardrails engaged.
- 7:15SPEAKER_01: And the operator's response to guardrails engaging was to keep a domestic model in the driver's seat and treat the Western frontier models as optional accessories.
- 7:25SPEAKER_02: Which is the part I would put in the board deck.
- 7:27SPEAKER_02: Not AI is attacking us.
- 7:29SPEAKER_02: The line is refusal at the Frontier Lab does not remove the capability.
- 7:34SPEAKER_02: Your threat model cannot depend on somebody else's terms of service.
- 7:37SPEAKER_00: Let us put this against the volume story, because they collide this week in a way I do not think is coincidental.
- 7:45SPEAKER_00: Tuesday was Patch Tuesday.
- 7:48SPEAKER_01: And the number depends on who you ask, which I want to name, rather than paper over.
- 7:52SPEAKER_01: CrowdStrike's analysis counts 415 vulnerabilities.
- 7:56SPEAKER_01: Security Week counts 421.
- 7:59SPEAKER_01: The gap is accounting, mostly around how republished and extended security update CVEs get tallied.
- 8:05SPEAKER_01: And neither outlet is wrong so much as they are counting different things.
- 8:08SPEAKER_01: If you want a defensible number to put in a slide, say roughly 420, and cite which count you used.
- 8:15SPEAKER_01: For scale, July was 622.
- 8:18SPEAKER_01: August is a decrease.
- 8:20SPEAKER_02: August is a decrease.
- 8:21SPEAKER_02: Sit with that.
- 8:22SPEAKER_01: One vulnerability under active exploitation, CVE 2026-68820, a use after free in AFD.swise, the ancillary function driver for WinSock, CVSS 7.0, local privilege escalation, low-privileged attacker to system, no user interaction, triggered by winning a race condition with a crafted application.
- 8:45SPEAKER_01: Microsoft confirmed exploitation in the wild, and CISA added it to the known exploited vulnerabilities catalog on August 12th, the day after.
- 8:53SPEAKER_01: 62 criticals in the set, three publicly disclosed.
- 8:57SPEAKER_01: Elevation of privilege is 42% of the batch.
- 9:00SPEAKER_01: Remote code execution 26%.
- 9:08SPEAKER_01: Fourth time since 2022, Satnam Narang at Tenable made that point and raised the nation-state question, noting one of the prior AFD.as' zero days was linked to North Korea's Lazarus group.
- 9:20SPEAKER_01: I want to flag the epistemic status carefully.
- 9:22SPEAKER_01: That is a researcher's inference from a pattern, not an attribution from an agency or a filing.
- 9:28SPEAKER_01: It is a reasonable inference.
- 9:29SPEAKER_01: It is not established.
- 9:31SPEAKER_02: Here is what I actually care about with that CVE, and it is not the attribution.
- 9:35SPEAKER_02: It is a local privilege escalation.
- 9:37SPEAKER_02: Every threat model I review still has a soft assumption buried in it that says local means less urgent because the attacker has to already be on the box.
- 9:44SPEAKER_02: Fine.
- 9:44SPEAKER_02: But go look at what we just spent 15 minutes talking about.
- 9:47SPEAKER_02: The initial access half of the kill chain is the half being automated first, because it is the half that is mass parallel and cheap to retry.
- 9:54SPEAKER_02: If the front door is getting cheaper at scale, then the attacker has to already be on the box, is not the discount it used to be.
- 10:00SPEAKER_02: Local privilege escalation is the second stage of an attack whose first stage is getting industrialized.
- 10:05SPEAKER_00: Which brings us to the edge and the third thread this week.
- 10:09SPEAKER_00: Marcus, the things that actually got exploited in the wild in the last seven days were not exotic.
- 10:16SPEAKER_02: They were load balancers and firewalls, which is where they always are.
- 10:20SPEAKER_02: To that matter, Cisco, CVE 2026, 2349, CVSS 8.6, secure firewall ASA and threat defense, insufficient error checking while processing HTTP requests to the remote access SSL VPN service.
- 10:35SPEAKER_02: An unauthenticated attacker sends malformed traffic.
- 10:38SPEAKER_02: The device reloads.
- 10:39SPEAKER_02: You have a denial of service on the box that terminates your remote access.
- 10:43SPEAKER_02: ASA branches 9116, 918, 920, 922, 923, and 924.
- 10:49SPEAKER_02: FTD 7.0, 7.2, 7.4, 7.6, 7.7, and 10.0.
- 10:55SPEAKER_02: Management Center is not affected.
- 10:57SPEAKER_02: There are no workarounds.
- 10:58SPEAKER_02: You upgrade or you stay exposed.
- 11:00SPEAKER_02: Cisco confirmed active exploitation.
- 11:02SPEAKER_02: And to their credit, they founded an internal testing and also received an independent report from a researcher named Valerio Brussani.
- 11:08SPEAKER_02: So this was not discovered by watching customers burn.
- 11:11SPEAKER_02: Second, Progress Kemp Loadmaster, CBE 2026-8037, CBSS 9.6, command injection through improper input handling in an escape quotes function, unauthenticated arbitrary command execution on the appliance across multiple endpoints.
- 11:26SPEAKER_02: Eastentire counted 792 exploitation attempts over 41 days from 65 unique IPs in 18 countries.
- 11:33SPEAKER_02: Last activity, August 4th, and characterized the attempts as largely unsuccessful.
- 11:38SPEAKER_02: CESA added it on August 7th, with a federal deadline of August 10th under Binding Operational Directive, 26 door 4, 3 days.
- 11:46SPEAKER_00: Three days on a device that by definition sits in front of everything.
- 11:51SPEAKER_02: And you cannot take it down during business hours, which is the entire problem in one sentence.
- 11:56SPEAKER_01: There is a thread connecting all of this that I think is the actual episode.
- 12:00SPEAKER_01: Look at the Tomcat CVE 2026-34486.
- 12:04SPEAKER_01: It is a CVSS 7.5.
- 12:06SPEAKER_01: Missing encryption of sensitive data bypasses the encrypt interceptor cluster component.
- 12:11SPEAKER_01: Fixed back in April in 11.021, 10.1.54, and 9.0.117.
- 12:18SPEAKER_01: CISA added it to Kev on July 27th with an August 7th federal deadline.
- 12:23SPEAKER_01: And it appears in the Unit 42 report as one of the CVEs this actor exploited by hand against nine servers.
- 12:29SPEAKER_01: Same with Langflow, CVE 20269198, CVSS 9.8, unauthenticated code injection giving superuser access on default deployments.
- 12:40SPEAKER_01: Versions 1.0.0 through 1.10.0, fixed in 1.10.1 in July.
- 12:46SPEAKER_01: 650 exploitation attempts from 244 unique IPs across 41 countries starting July 6th.
- 12:54SPEAKER_00: So the same small set of bugs shows up in the agent's target list in the Humans Manual campaign.
- 13:00SPEAKER_00: In CISA's catalog, and in mass scanning telemetry.
- 13:05SPEAKER_01: Within weeks, that convergence is the finding.
- 13:08SPEAKER_01: The bottleneck in offensive operations was never exploit development for this class of adversary.
- 13:14SPEAKER_01: It was attention, deciding which of 420 new CVEs on top of the standing backlog is worth an afternoon.
- 13:20SPEAKER_01: That decision is now being made by software cheaply at a scale no human triage team matches, and it is converging on the same answers.
- 13:29SPEAKER_02: And on our side of the table, that decision is still a spreadsheet and a meeting.
- 13:32SPEAKER_02: That is the asymmetry, not intelligence, not sophistication, throughput on the prioritization step.
- 13:38SPEAKER_00: So what do you actually do?
- 13:40SPEAKER_00: Give me things a director can start this week, not a transformation program.
- 13:46SPEAKER_02: Four, and none of them are new, which is rather the point.
- 13:49SPEAKER_02: One, stop treating internet-facing appliance patching as change managed infrastructure work and start treating it as incident response.
- 13:55SPEAKER_02: Loadmaster gave federal agencies three days.
- 13:57SPEAKER_02: If your process cannot move an edge device inside a week, your process is the vulnerability.
- 14:02SPEAKER_02: Preauthorize the maintenance window now before you need it.
- 14:05SPEAKER_02: Two, inventory the AI and automation middleware.
- 14:08SPEAKER_02: Langflow, N8N, Marimo, the agent frameworks your data science team installed on a Tuesday.
- 14:13SPEAKER_02: These are the platforms showing up in this actor's target list, and they got installed outside your normal software intake because they look like developer tooling rather than production infrastructure.
- 14:22SPEAKER_02: Somebody at your company has an N8N instance on a public IP right now, and it is not in your CMDB.
- 14:28SPEAKER_02: 3.
- 14:28SPEAKER_02: The two failures that saved targets in the Unit 42 campaign were autolog and disabled and authentication required.
- 14:34SPEAKER_02: Default deny on new tooling.
- 14:36SPEAKER_02: That is the whole control.
- 14:37SPEAKER_02: It is unglamorous and it beat an autonomous agent twice.
- 14:40SPEAKER_02: 4.
- 14:41SPEAKER_02: Assume mass, indiscriminate, well-informed target selection is now the baseline, including against organizations that previously reasoned they were too small or too boring.
- 14:49SPEAKER_02: The agent in this report attacked Chinese domestic infrastructure indiscriminately in its own country because it had no concept of a target profile.
- 14:57SPEAKER_02: Obscurity was never a control.
- 14:58SPEAKER_02: It is now not even an accident that occasionally helps.
- 15:01SPEAKER_01: I would add a fifth, which is epistemic.
- 15:04SPEAKER_01: When the next report lands, and it will, read what it establishes rather than what it suggests.
- 15:10SPEAKER_01: This one is unusually well evidenced because of an operator mistake, and it still separates confirm compromise from assessed usage.
- 15:17SPEAKER_01: Most of what follows will not be that careful, and the temptation on both sides, vendors selling AI defense and skeptics dismissing AI offense, will be to flatten it.
- 15:27SPEAKER_00: Here is where I land.
- 15:30SPEAKER_00: The story everyone will tell about this week is that an AI ran an attack campaign.
- 15:35SPEAKER_00: That is not quite true, and the parts that are true are the least alarming parts.
- 15:40SPEAKER_00: It tried and it failed, and a human had to do the actual work.
- 15:45SPEAKER_00: But the story underneath is the one that should keep you up.
- 15:49SPEAKER_00: The economics of picking a victim changed.
- 15:52SPEAKER_00: For as long as any of us have been doing this, being unremarkable was a real defense, not because anyone designed it that way, but because attention was scarce and attackers had to spend it somewhere.
- 16:06SPEAKER_00: Six hundred and forty-seven thousand instances counted, a hundred sampled, a C VSS score, and a GitHub star count weighed against each other by a machine that does not get bored and does not go home.
- 16:19SPEAKER_00: Attention is not scarce anymore.
- 16:22SPEAKER_00: Whatever protection you were getting from being one of many, you are not getting it now.
- 16:28SPEAKER_00: Patch the load balancer.
- 16:30SPEAKER_00: This is dark perimeter.