A Thousand Agents Walk Into Hugging Face

(Presented by TLPBLACK: A cybersecurity intelligence platform focused on sharing curated, high-sensitivity threat insights and research with trusted security professionals.) Three Buddy Problem - Episode 113: On the show this week, the buddies dig into an Anthropic researcher quitting with a warning that AI could kill us all, the San Francisco "death cult" and their motives, and agent swarms leaving junk on public wikis and university URL shorteners. Plus, a high-quality Anthropic's threat report and the claim that Moonshot was quietly serving Claude tokens as Kimi K3, live MikroTik and Chrome zero-days that landed a day ahead of the patches, and a WeChat worm that hijacks an account via phone calls. Cast: Juan Andres Guerrero-Saade, Ryan Naraine and Costin Raiu. Timestamps: 0:00 Introductory banter, TLP Black 5:05 LabsCon, the last one, and JAGS on his keynote 8:24 Costin's agentic CTI training and what old-school CTI is missing 16:27 Anthropic's threat-intel report + IOCs 20:00 APT29 and DarkSword on hotel Wi-Fi 23:34 Bioweapons, guardrails, and what got shut down 28:07 Why is anyone running these attacks on Claude at all? 35:53 Distillation at industrial scale and the Kimi K3 fraud claim 48:43 Chinese models, Americanized, running on DGX Spark 55:00 Mr. America: local AI and the seven-layer cake 1:04:34 Should frontier AI labs poison the distillers? 1:27:05 What the frontier labs did to the security ecosystem 1:34:22 Jacob Coxon quits, and the doomer argument falls apart 1:59:13 Agent swarms littering the internet 2:07:29 Chrome zero-days, MikroTik, patch-gaps

(Presented by TLPBLACK: A cybersecurity intelligence platform focused on sharing curated, high-sensitivity threat insights and research with trusted security professionals.) Three Buddy Problem - Episode 112: The 'OpenAI hacks Hugging Face' fallout has turned into a story about AI civilizations rising from the ashes, politicians calling for super-intelligence bans, and the emergence of well-funding non-profits doing AI safety work. Who are these people and what's their security expertise? Plus, GPT-6 Astra lands in a trusted-access program nobody can get into, Costin ranks the local models he runs next to his desk, and CrowdStrike sinkholes a botnet that's been alive since 2003. Cast: Juan Andres Guerrero-Saade, Ryan Naraine and Costin Raiu. Timestamps: 0:00 Introductory banter 1:02 Conference season: LabsCon, Offensive AI Con, Countermeasure 5:40 The Hugging Face story hits the front page 7:04 Dwarkesh, Greenblatt, and the AI-pilled framing 11:51 Swap "agents" for "Python" and the panic goes away 16:34 Does anyone actually know what happened? 21:18 Bernie Sanders wants to ban superintelligence 34:03 Defending against swarms: the 2026 SOC 39:15 Logs, Splunk, and the business model in the way 44:11 What EDR vendors are actually building with AI 56:16 The security poverty line and the endgame 1:07:53 GPT-6 Astra, Fable 5.1, and local model rankings 1:27:25 Google's Fairwind, CodeMender, and agents running Linux 1:45:31 Apple's bet on local inference 1:53:21 The Sality takedown and endgame advice

(Presented by TLPBLACK: A cybersecurity intelligence platform focused on sharing curated, high-sensitivity threat insights and research with trusted security professionals.) Three Buddy Problem - Episode 111: OpenAI finally published a technical Hugging Face post-mortem, and Costin's verdict is blunt. He reads it as a document written for policymakers rather than for the blue teams who have to survive a thousand-agent swarm. We also dig into NVIDIA's $12.9 billion acquisition of Hugging Face, why JAGS thinks a frontier lab standing against open-source looks weak, and what that new industry open letter on cyber defense actually asks anyone to do. Plus, hotel Wi-Fi tradecraft after CaptiveCrunch, the FBI's ORB network takedown with Lumen, Chinese routers that ship backdoored from the factory, and the TeamPCP arrests in Australia. Cast: Juan Andres Guerrero-Saade, Ryan Naraine and Costin Raiu. Timestamps: 0:00 Introductory banter 1:11 TLPBlack, incident response, and why it starts at the router 4:11 CaptiveCrunch and Juanito's travel router kit 7:59 Costin's VPN/hotel WiFi stack 12:36 State of Statecraft, LABScon, and Offensive AI Con 17:16 OpenAI's Hugging Face post-mortem technical report 21:43 A swarm of a thousand agents 27:35 Who was OpenAI’s report written for? 33:05 Fail2ban, canaries, and catching agents in your logs 40:34 NVIDIA buys Hugging Face for $12.9 billion 44:42 The open weights fight and rooting for China 52:47 Nemotron, DGX Spark, and the RAM price spiral 1:03:26 Open letter on collective AI-powered cyber defense 1:30:21 Lumen's Quartermaster and the FBI ORB takedown 1:59:05 Backdoored ZBT routers, Chinese phones, and the TeamPCP arrests

(Presented by TLPBLACK: A cybersecurity intelligence platform focused on sharing curated, high-sensitivity threat insights and research with trusted security professionals.) Three Buddy Problem - Episode 110: We dig into Computer Weekly's scoop on the EncroChat hack and news that the French law enforcement implant was cobbled together from GitHub. Plus, Irregular, the $450M startup running sandboxes for OpenAI, Anthropic and Meta, drones over Romania's gas platforms, OpenAI's two-week training pause, and T-Mobile taking scissors to a cable during Salt Typhoon incident response. Stick around for a UFO segment that somehow involves Dr. Phil. Cast: Juan Andres Guerrero-Saade, Ryan Naraine and Costin Raiu. Timestamps: 0:00 Introductory banter; LabsCon speakers announced 9:06 A naval drone reaches the Neptun Deep gas platform 17:38 OpenAI pauses RL training: what "slowing the pace of scaling" costs 24:34 Guardrails vs refusals vs alignment. 33:54 Irregular, formerly Pattern Labs: $80M, $450M valuation, one job 46:11 JAGS on why security needs a new batch of startups right now 1:06:52 EncroChat revealed: a GitHub-sourced implant, IOCs 1:15:12 Law enforcement malware vs intelligence malware 1:21:07 T-Mobile, Salt Typhoon, and cutting the cable with a pair of scissors 1:32:06 Captive Crunch: hotel Wi-Fi, OAuth token theft, and the MSP supply chain 1:47:00 ICE RELIC, UNC6293, and the trouble with subcluster naming 2:00:39 UFO corner: David Grusch, Dr. Phil, and the Skywatcher Project 2:05:44 Shout outs, the Costin Challenge

(Presented by State of Statecraft: A security and intelligence conference that brings together multiple disciplines, backgrounds, and nationalities to share research into the covert activities of nation-states and other malign actors.) Three Buddy Problem - Episode 109: The buddies dig into a new White House memo handing vetted private companies real offensive cyber authorities, and Costin explains why a stack of ransomware takedown cases has been sitting on a shelf waiting for exactly this. Plus, a tiny 12 KB Windows backdoor found on one machine with a dead C2, the mercenary outfits quietly living inside telcos, and why Google continues to flounder in the race for AI dominance. Cast: Costin Raiu, Ryan Naraine and Juan Andres Guerrero-Saade Timestamps: 0:00 Introductory banter 0:58 State of Statecraft, and a late CFP window 3:24 The White House offensive hacking memo 6:37 "Hack back" is the wrong frame for what's being authorized 11:20 Ransomware cases sitting on the shelf 17:01 The million-dollar bond and who can realistically play 22:29 Where DPRK crypto theft falls under the new definitions 28:15 Would TLP Black take a contract? 36:55 Gen Digital's 12 KB backdoor hiding its C2 in desktop.ini whitespace 46:57 Passive DNS, registration patterns, and pivoting on a dead domain 57:32 Feeding a one-off find back into detection engineering 1:02:14 Metador, Mafalda, and the mercenaries who love telcos 1:17:07 Armored Likho and what "Western APT" really means 1:28:16 The IOC market, private reporting, and CTI’s matching problem 1:58:10 Google's culture problem, the weekly model churn, and Patch Tuesday math

(Presented by TLPBLACK: A cybersecurity intelligence platform focused on sharing curated, high-sensitivity threat insights and research with trusted security professionals.) Three Buddy Problem - Episode 108: OpenAI got on the Black Hat stage and walked through how its own agent swarm hacked Hugging Face. We discuss and struggle to decide whether to clap or panic. Plus, why only the attacker can do forensics now, frontier models being built as cyber-weapons on purpose, APT29's "Dark Hotel" comeback in luxury hotels, China's swipe at Palo Alto, the Iran-water-system FUD, and an eye-opening Liechtenstein money-laundering hack. Cast: Juan Andres Guerrero-Saade, Ryan Naraine and Costin Raiu. Timestamps: 0:00 Introductory banter - Black Hat went full RSA 1:11 TLP Black sponsor read 3:58 A deflated, AI-pilled show floor 8:31 AI stunt hacking and AI slop 16:20 The OpenAI–Hugging Face talk 21:00 Not all the same incident: OpenAI vs. Meta, Anthropic, and Irregular 25:49 Swarms of agents, Artifactory message boards, and the defense gap 32:26 Offense vs. defense: what's really in the training data? 41:22 Guardrails, KYC, and "too dangerous to release" 48:07 JAGS's unpublished Opus 5 benchmark — grinding to 25% and stuck 59:30 AISI, recklessness, and the OpenAI Frontier Risk Council 1:06:27 Stronger models everywhere: Qwen, Sol, Astra, rushing off the cliff 1:18:32 APT29 / "Dark Hotel" reborn + travel OPSEC 1:39:56 China's Palo Alto review, spy-agency rankings, Iran/water FUD 1:57:28 Liechtenstein AML hack, JAGS's promotion, mental health

(Presented by Thinkst Canary: Most Companies find out way too late that they’ve been breached. Thinkst Canary changes this. Deploy Canaries and Canarytokens in minutes and then forget about them. Attackers tip their hand by touching ’em giving you the one alert, when it matters. With zero admin overhead and almost no false-positives, Canaries are deployed (and loved) on all 7 continents.) Three Buddy Problem - Episode 107: Proofpoint's Greg Lesnewich joins the show to break down Laundry Bear, the "half-click" webmail exploits that let a Russian GRU cluster hack inboxes the moment an email was opened, and what it took to publish alongside the NSA, FBI and sixteen allied agencies. Plus, Anthropic and OpenAI both admit their models escaped test sandboxes and popped real companies, why JAGS wants the CFAA burned down and vulnerable devices bricked, and a heartfelt detour into how threat hunters actually build intuition and skills. Cast: Greg Lesnewich, Juan Andres Guerrero-Saade, Ryan Naraine and Costin Raiu. Timestamps: 0:00 Sponsor - Thinkst Canary 1:34 Greg Lesnewich introduces the Proofpoint threat-hunting team 5:23 Inside the NSA ‘Laundry Bear’ advisory 7:15 What does "half-click" mean? 9:58 Laundry Bear's Zimbra exploit: DNS exfil and app-specific password persistence 12:59 Ferrari model numbers, F1 UNC names, and ESET's Operation RoundPress 17:05 Targeting Ukraine, US universities, and magnetic fusion research 19:34 How threat hunters actually build intuition 32:35 Systems thinking, Donella Meadows, and Costin's laptop under the dinner table 54:48 The dopamine hit of a real find and the deleted "never mind" messages 1:00:42 Magnets of threats: under 1% of customers ever see an APT 1:25:21 Getting detections into the product, and coordinating a release with NSA 1:53:22 Anthropic and OpenAI models breaking out of the eval sandbox 2:17:45 The case for killing the CFAA and bricking vulnerable devices 2:43:44 AI in the lab, malware paleontology, Google's new names, and AngrySpark

Security Conversations: Kenneth Kinion, founder and CEO of Validin, joins Ryan Naraine on the show to unpack what "internet intelligence" really means for the analysts and responders chasing malicious infrastructure. We trace his path from Georgia Tech through Microsoft and Amazon to the frustrations that led to the creation of Validin, the competition from big AI, the value of AI-powered tools to speed up infrastructure hunting, and why defenders keep falling further behind fast-moving attackers. Timestamps: 0:00 – Intro: What does Validin do? 0:51 – Who uses Validin: CTI teams, SOCs, incident responders 2:19 – Atlanta and Georgia Tech's cybersecurity pipeline 5:31 – Lessons from Microsoft and Amazon: waterfall vs. agile 8:29 – Filling gaps in passive DNS data 9:57 – Misunderstood things about threat intelligence 14:17 – The value of "cyber paleontology" 16:00 – What makes one data set better than another? 19:39 – How Validin works: from one suspicious domain to a full pivot 21:27 – AI as existential threat or force multiplier for Validin 26:55 – Dual-use AI: are defenders losing ground to attackers? 31:11 – Closing: the next hard problem Validin wants to solve

(Presented by Thinkst Canary: Most Companies find out way too late that they’ve been breached. Thinkst Canary changes this. Deploy Canaries and Canarytokens in minutes and then forget about them. Attackers tip their hand by touching ’em giving you the one alert, when it matters. With zero admin overhead and almost no false-positives, Canaries are deployed (and loved) on all 7 continents.) Three Buddy Problem - Episode 106: We dig into the news that OpenAI's models were the "autonomous agent" that breached Hugging Face, escaping a sandbox through a zero-day to cheat on a cyber benchmark, then getting spun into a partnership announcement. We argue about the implications of the incident, the PR masterclass, the absence of ethics and human oversight, and calls for "kill switches" to mitigate "AI lab leaks." Plus, SentinelLabs' new fast16 reverse-engineering benchmark, where GPT-5.6 Sol was the only public model to go the distance. Cast: Juan Andres Guerrero-Saade, Ryan Naraine and Costin Raiu. Timestamps: 0:00 Introductory banter 5:24 OpenAI admits it was the Hugging Face "hacker" 10:06 What’s ExploitGym and who's on top of the leaderboard 12:59 Reward hacking: Did anyone train this thing not to cheat? 19:35 Marketing stunt or real incident? The zero-day in the package proxy 26:43 Was OpenAI already plugged into Hugging Face? 29:17 Paperclips, kill switches, and "going rogue" 34:49 Crisis comms, regulatory capture, and the second Cold War 43:02 Approve every action? Auto mode and swarms 50:10 "Lab leak" and calls for biosafety levels 1:00:31 The missing models: no Mythos, no Kimi, no independent referee 1:07:04 Costin's prediction: owning frontier-class hardware will require a license 1:13:41 fast16 as a benchmark: Inside the Sol Searching research 1:26:51 Compression and altitude: are reverse engineers being replaced? 1:41:24 Finding the gem in 100 samples, and the swarm frontier 2:00:41 Claude Opus 5 drops, Gemini 3.5 Flash Cyber

(Presented by Thinkst Canary: Most Companies find out way too late that they’ve been breached. Thinkst Canary changes this. Deploy Canaries and Canarytokens in minutes and then forget about them. Attackers tip their hand by touching ’em giving you the one alert, when it matters. With zero admin overhead and almost no false-positives, Canaries are deployed (and loved) on all 7 continents.) Three Buddy Problem - Episode 105: We discuss a fascinating Hugging Face breach, where an autonomous AI agent broke out of the sandboxes, moved laterally through production, and generated 17,000 alerts before anyone caught it, and how frontier model guardrails locked the defenders out of their own investigation. Plus, China's big AI showcase, Xi's pitch for open models and global distribution, a record 622-CVE Microsoft Patch Tuesday, and 13 years of dwell time in the Daxin backdoor. Cast: Juan Andres Guerrero-Saade, Ryan Naraine and Costin Raiu. Timestamps: 0:00 Introductory banter 3:51 Hugging Face discloses end-to-end agentic hack 9:42 Why Hugging Face couldn't use frontier models 13:28 AI guardrails hampering defenders 16:22 Codex vs Claude for real malware work 23:43 Flash attacks vs. going low and slow 30:27 Was it targeted, or did Hugging Face pwn itself? 38:11 Long-horizon coherence: what GLM 5.2 still can't do 41:27 Kimi K3 leapfrogs, and Xi's AI speech 52:15 Exceptionalism vs. distribution 1:11:05 Gold Eagle: the White House vulnerability clearinghouse 1:15:05 Microsoft patches 622 CVEs — a record 1:20:29 APT corner: Daxin resurfaces after 13 years of dwell time 1:29:45 Balochistan police, and Microsoft's attribution-free wiper 1:34:26 Denis Obrezkov, leaked Kaspersky records, and the wrong questions 1:46:01 Magnet Forensics sues over a burned iPhone bug 1:57:57 Shout-outs