Dark Perimeter: Real Breaches, Real Stakes

Cole Drayden

Every major cyberattack has a story behind it. A vulnerability no one patched. A phishing email someone clicked. A nation-state with a motive. Dark Perimeter goes beyond the headlines to explore the true stories of the hacks, breaches, and cyber operations that shaped history - told in narrative form for security professionals and curious minds alike. No guests, no panels, no filler. Just the story.

  1. 6d ago

    Ground Stop, Part 1: Extraordinary Circumstances

    On Friday, November 13, 2026, at 3:54 p.m. Eastern, the FAA ordered every airport in the United States to hold its departures. The stop lasted four hours and twenty-one minutes, and the internet filled the silence with cyberattacks, rogue AI, a staged crisis and aliens. In part one of a new series, Cole Drayden gives the verdict first, then sits down with two researchers who spent the weekend in the record, June Calloway and Eli Navarro, to reconstruct the forty-five minutes before the order: the first ground stops in Chicago, a controller in Fresno working around a bad frequency, the spread from center to center, the wording of the advisory, and what the FAA says broke. No theories tonight. Those come later, and they get tested. Disclosure. June Calloway and Eli Navarro are fictional composite characters. They stand in for the research behind this series and claim no aviation credentials. Facts come from the sources below. We name agencies, companies and public officials, and we do not name private individuals. The verdict so far (our read, analysis and not sourced fact). This was very likely an accident: a network failure set off by routine contractor work. We found nothing that supports aliens or a deliberate clearing of the skies, and several facts cut against both. The official account is still unfinished. The root cause has not been published, and the explanation changed more than once in its first twenty hours. How we label claims. Documented: from a primary record. Reported: from a named outlet. Inferred: our own reasoning. Speculation: said out loud as speculation. Not yet verified by us. We have read the reporting on the controller audio. We have not yet pulled the recordings and listened ourselves, so everything from that audio is labeled reported. Sources. The official account, times, aircraft count, centers affected, cancellations and the router explanation: FAA, Nationwide Ground Stop: Questions and Answers (documented) Minute by minute timeline, the spread from Chicago, the 3:44 p.m. first ground stop, the White House briefing and the funding request: The Air Current (reported) Fresno controller audio before the order: ABC30 Fresno (reported) Chicago ground stops at about 3:10 p.m. and Chicago Center radio traffic: Glitchwire (reported) Longest nationwide halt since September 11, 2001: Reuters, via Yahoo Finance (reported) Controllers' union and FAA administrator remarks, and the Amarillo timing: Yahoo News live coverage (reported) The Transportation Secretary on the Amarillo site and the unknown root cause: CBS News, Face the Nation transcript, November 15 (documented) How the official explanation shifted in its first twenty hours: GovInfoSecurity (reported) Starship launch time: Starship flight 15 launch record (secondary) Europe holding flights at the FAA's request, and diversions as gates filled: Wikipedia, 2026 FAA nationwide ground stop (secondary, citing USA Today and The New York Times) Fake footage already circulating: Lead Stories, via Yahoo News (reported) Corrections. None so far. If we got something wrong, the correction goes here. Next in the series: how one router at a remote radio site reached seventeen control centers. Support the show

  2. Nov 7

    Two Sentences

    On October 27, two sentences pushed from a server changed what Claude Code will build: no more recreations of specific existing games, even renamed. Cole Drayden, Dr. Elliott Vance and Marcus Hale look at where agentic AI really stands in late 2026, why the rule appeared after a month of AI built browser ports of commercial games, and the unsettled law of AI clean room rewrites. Then the options people reach for when a sanctioned tool says no (DeepSeek, hosted open weights, rented GPUs, fully offline models), what each costs and risks, and five controls security teams should put in place before the workarounds start. Disclosure. Dr. Elliott Vance and Marcus Hale are fictional composite characters. Facts and figures come from the sources below; the items under "our read" are analysis, not reporting. Nothing in this episode is legal advice. Our read (analysis, not sourced fact). Guardrails in sanctioned tools push work toward unsanctioned ones, so a refusal is a security signal. Small local models are useful for private, bounded tasks but fall short of frontier agents on long jobs. Renting a multi GPU node can cost more than API tokens unless agents run near continuously. Vendor model policy is a runtime dependency that belongs in third party risk tracking. Sources. Claude Code game rule: How2Shout, Engadget, Gizmodo, VideoCardz Clean room law: Lexology, The Clean Room Problem, Hexaware, Jia Wei Ng Adoption data: Camunda via Techstrong, Deloitte, McKinsey, Forrester DeepSeek: V4.1 Flash, Super Agent (Dataconomy), Super Agent (The Next Web), DHS exclusion order summary GPU pricing: Neurohive, October 2026, provider comparison Support the show

  3. Nov 6

    It Passed Every Check

    On October 21, the release pipeline of LiteLLM, a widely used open source gateway between applications and large language model providers, built, signed and published a poisoned release that stole cloud, SSH, browser and provider credentials from any machine where Python ran. It was live on PyPI for one hour and forty-three minutes and was downloaded roughly 49,000 times, and stolen cloud credentials were being tested against model inference services about four and a half hours later. Cole Drayden, Dr. Elliott Vance and Marcus Hale walk the victim's own timeline: an administrative key committed to a public repository in May and never revoked, an organization owner's GitHub CLI token handed back by an internal agent platform on October 10, branch protection overridden five seconds after it fired, a lookalike organization, and three stolen identities with three distinct roles. Then the honest balance of what is and is not known, six changes a security program should make, and where identity as the supply chain goes next. Specifics kept out of the audio. Package: litellm 1.107.0 (PyPI, all seven wheels, live 11:39 to 13:23 UTC on 2026-10-21). Container tags v1.107.0, latest, main-latest and main-stable on Docker Hub and GHCR, cleaned at 13:51 and 14:20 UTC. Lookalike organization: Berri-AI/litellm (the real one is BerriAI/litellm). Payload files dropped in the site-packages root: cook.pth, cook.py, sitecustomize.py. Exfiltration host: a lookalike domain, www.litellm.dev. Last safe version per LiteLLM and Wiz: 1.106.2. Key dates: admin key committed 2026-05-27; owner GitHub CLI OAuth token obtained 2026-10-10; secrets exfiltrated 2026-10-10 and 2026-10-12; branch protection override 10:08:34 UTC on 2026-10-21; community issue filed 12:06 UTC; Wiz observed credential abuse 16:20 to 17:01 UTC; LiteLLM root cause report 2026-10-23; LiteLLM "What we are changing after October 21" 2026-10-30. March 2026 precedent: a compromised Trivy dependency, unpinned in CI, read static registry credentials from environment variables. Opening reference: the Binance theft of 2026-10-29, traced by TRM Labs (about $6.36 billion) and Elliptic (about $6.43 billion), with FBI attribution to North Korea's TraderTraitor reported on 2026-10-31. Primary sources. LiteLLM incident post: docs.litellm.ai/blog/security-incident-oct-21-2026. LiteLLM attack chain report: docs.litellm.ai/blog/security-incident-oct-21-2026-attack-chain. LiteLLM changes: docs.litellm.ai/blog/security-changes-after-oct-21-2026. Microsoft Security: Inside the LiteLLM 1.107.0 supply chain compromise. Wiz Research: LiteLLM compromised, steals credentials. StepSecurity: LiteLLM 1.107.0 compromised. TRM Labs and Elliptic on the Binance theft: TRM Labs, Elliptic. Not legal advice. No ads. No sponsors. This is Dark Perimeter. Support the show

  4. Nov 4

    Fifteen Minutes to Root

    A repository whose entire product is a prompt crossed 41,000 stars in under ten days, and what it does is strip the advertising, the telemetry and the stock launcher out of your television in about fifteen minutes. In the same month, a three-flaw chain was published that takes root on an LG set from the local network with no PIN and no action at all from the owner. Cole Drayden, Dr. Elliott Vance and Marcus Hale work both directions of the same story: what the debloating movement actually removes, what automatic content recognition was doing to justify it, the attack surface you create by leaving the debug bridge open, and what a building full of unmanaged televisions means for a real security program. Specifics kept out of the audio: the owner-side projects are cleanyourtv (Gathin23/cleanyourtv) and dumbtv (humblemuzzu/dumbtv). The Samsung Tizen root work chains SVE-2025-50109 for the initial foothold with a physical-memory bug on QN90B-generation panels and a Mali GPU bug (CVE-2025-0072) on QN90F. The webOS chain was shown at TyphoonPWN 2026 and published by SSD Secure Disclosure on 12 October 2026, tested on the 32LQ636BPSA and 27TQ600SW on webOS 7.5.7-25; no CVE identifiers were assigned and LG has shipped a fix. The ACR research is Vekaria et al., ACM Internet Measurement Conference 2024 (UCL / UC Davis), with related work from UC3M on data collection during external-display use. The ADB-exposure botnets referenced are Kimwolf (reported by QiAnXin XLab, with analysis from Synthient) and XCPE, the 1.3 million device Videostrong campaign Google disrupted. No ads. No sponsors. This is Dark Perimeter. Support the show

  5. Oct 8

    Critical Is Not a Priority

    Four things carried a critical severity rating this week, and they were not the same problem. Cole, Dr. Vance and Marcus Hale rank the week out loud, with reasons, because the ranking is the actual work. Cisco Catalyst SD-WAN Manager (CVE-2026-76504), CVSS 9.8. A URL-encoding authentication bypass yielding admin API access. Cisco confirmed exploitation in the wild, learned of it through a TAC support case, and the KEV remediation date of 3 October has already passed. FortiMail (CVE-2026-104286), CVSS 9.8. Unauthenticated arbitrary file write. Exploitation acknowledged by Fortinet, published indicators include an ld.so.preload entry, and for some branches the workaround is the remediation rather than a bridge to one. Atlassian (CVE-2026-21589), CVSS 9.3. Unauthenticated file access across eight Data Center products, constrained unless a Crowd deployment leaves credentials in a guessable path. Attempts reported against honeypots; no confirmed compromise as of recording. Chrome 155, 247 fixes and four critical use-after-free bugs, with no stated in-the-wild exploitation. Volume is not urgency. Then the week's largest event by population affected: approximately 8.8 million people in Denmark's national population register, reached through a private company's lawful lookup access over roughly ten days in September. No exploit, no severity score, no patch. Throughout, we separate what a vendor confirmed from what a researcher reported from what a secondary source inferred, and we name two places where the public record conflicts with itself rather than quietly picking a side. Support the show

  6. Oct 6

    Your Tier Zero List Is Wrong

    Three things happened in the last week that look like three different stories. Citrix shipped an emergency bulletin for NetScaler bugs that were already being exploited. GitLab patched a 9.9 in the gateway that runs its AI agents. Dell pushed six critical fixes for the component connecting enterprise storage arrays to Kubernetes. Cole Drayden, Dr. Elliott Vance and Marcus Hale argue they are one story. On NetScaler: CVE-2026-88771, an unauthenticated command injection, and CVE-2026-88772, a DTLS memory overflow, both CVSS v4 9.5, in the catalog the same day Citrix published. Exploitation attempts reported on the 26th of September with no CVE-specific detections available, customers told to disconnect on the 28th, and the bulletin on the 29th. Two different exposure counts from two different research teams, named rather than reconciled. The web shell tradecraft, the persistence in web-accessible directories, and the sentence that should change your runbook: patching will not remove access for an attacker who already has it. Plus the separate CVE-2026-88779 and why denial of service is a smaller claim than remote code execution. On GitLab: CVE-2026-90970, a prompt template sandbox escape in Duo Agent Platform custom flows, authenticated, self-hosted gateways only, no confirmed exploitation, and still the most interesting item of the week. A boundary between content and execution where the content is authored as configuration and never gets a program's review. On Dell Container Storage Modules: six criticals read as a set, ending in cluster-wide reads of Kubernetes Secrets and token forgery, which is a credential harvest with a storage CVE number attached. The thesis: in all three cases the compromised thing was not where the data lives, it was the thing that administers where the data lives. Ask your team for the tier zero asset list and you will get domain controllers. You will not get the VPN appliance, the CI/CD system or the storage driver that can read every Secret in the cluster. Tiering models ask what is stored here. The question that predicts this week is what can this reach, and with whose authority. Closes with four things to do this week, the first of which is not patching. Support the show

About

Every major cyberattack has a story behind it. A vulnerability no one patched. A phishing email someone clicked. A nation-state with a motive. Dark Perimeter goes beyond the headlines to explore the true stories of the hacks, breaches, and cyber operations that shaped history - told in narrative form for security professionals and curious minds alike. No guests, no panels, no filler. Just the story.