Podcast
Three Buddy Problem
Hosted by Security Conversations · EN
The Three Buddy Problem is a popular Security Conversations podcast that goes beyond industry talking points to discuss what others won’t -- nation-state malware, attribution, cyberwar, ethics, privacy, and the messy realities of securing computers and corporate networks. Hosted by three veteran security pros -- journalist Ryan Naraine and malware paleontologists Costin Raiu and Juan Andres Guerrero-Saade -- the weekly show attracts a highly engaged audience of security researchers, corporate defenders, CISOs, and policymakers. Connect with Ryan on Twitter (Open DMs).
47episodes
Episodes
Newest firstAll episodes
OpenAI's models breached Hugging Face, reward hacking ethics, benchmarking fast16
6d ago02:16:03Tap to summarize(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 CyberLinks:Transcript Sam Altman: "We had a significant security incident" OpenAI and Hugging Face partner to address security incident during model evaluation ExploitGym: Can AI Agents Turn Security Vulnerabilities into Real Attacks? US politicians float 'AI Kill Switch' US lawmakers push for AI 'kill switch' after OpenAI models go rogue OpenAI Daybreak Jensen Huang debuts on X Jensen Huang: Open Weights and American AI Leadership Can Frontier Models Tackle Autonomous Long-Horizon Malware Analysis? fast16 — IDA Databases and Analysis Artifacts Kimi K3 - API Platform Introducing Gemini 3.5 Flash Cyber NSA and Partners Alert Zimbra Collaboration Suite Users of a Russian State-Supported Phishing Campaign Russian State-Supported Cyber Actors Conduct Phishing Campaign Targeting Users of Zimbra Collaboration Suite Proofpoint: TA488 Targets Zimbra Mailservers with Half-Click Exploits CISA: Iran Cyber Actors Exploit PLCs Across US Critical Infrastructure Iran War Cyber Threat Landscape - A Midyear Assessment Thinkst Canary
Transcribe →
Hugging Face Just Got Hit by the First Fully Autonomous AI Attack
1w ago02:07:29Tap to summarize(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-outsLinks:Transcript Thinkst Canary Huggingface security incident disclosure White House Launches Gold Eagle Initiative for Vulnerability Coordination Trump admin unveils AI-supported clearinghouse for vulnerabilities YouTube: Xi Jinping speaks at world AI conference Microsoft Patch Tuesday by the numbers Rival Espionage Actors Converge On Pakistani Police Daxin Returns: Stealthy Malware Resurfaces in Taiwan GigaWiper: Anatomy of a destructive backdoor assembled from multiple malware IPhone Hacking Firm Sues Ex-Worker Over Alleged Theft of Secrets Paradigm Shift - Introducing usbliter8 LABScon 2026 TLPBLACK
Transcribe →
Microsoft's Secret Weapon: The GDID That Caught 'Scattered Spider' Teen
4w ago01:36:03Tap to summarize(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 104: We discuss the return of Anthropic's Fable 5 from export-control suspension with guardrails so aggressive that spelling "exploit" gets you downgraded. Plus, a debate on AI frontier labs killing businesses at scale, and OpenAI offering equity to the US government. Also, buried on page nine of a 'Scattered Spider' arrest indictment: Microsoft's never-before-detailed GDID device identifier, a persistent Windows fingerprint with massive implications for OPSEC, privacy, and APT tracking. Cast: Juan Andres Guerrero-Saade, Ryan Naraine and Costin Raiu. Timestamps: 0:00 Cold open: Heat wave in Washington DC 3:45 Fable 5 returns after the 15-day timeout 5:21 "Refined classifiers" and the downgrade-to-Opus mess 8:23 Codex vs. Claude: real-world malware analysis test 12:41 Who are the guardrails for? Defenders locked out 19:13 What even is a "jailbreak assessment framework"? 21:37 Two theories: failed PR vs. killing a thousand startups 24:59 Could the labs build kernels or a whole OS? 31:38 Bureaucracy is the moat 36:09 Can AI actually run an attack? (Spoiler: 14 detections) 47:01 OpenAI offers the US government a 5% stake 58:16 Scattered Spider arrest and Microsoft's GDID revelation 1:12:02 OPSEC fallout: how APT groups adapt to device telemetry 1:27:18 UFO update, shout-outs from SeoulLinks:Transcript Redeploying Claude Fable 5 (Anthropic) Howard Lutnick on Anthropic export controls OpenAI proposes handing Trump administration 5% stake DOJ: Alleged Member of “Scattered Spider” Extradited to USA Full DOJ complaint on Scattered Spider arrest HBO Max -- Q: Into the Storm Q Into the Storm LABScon 2026 Thinkst Canary
Transcribe →
US Gov Takes the Wheel: Who Gets to Use the Best AI?
Jun 2901:53:54Tap to summarize(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 103: We dive into the U.S. government's takeover of frontier-model rollouts (Mythos, Fable, and OpenAI's Sol/Terra/Luna) and what it means when intelligence gets commoditized but access gets rationed. Plus, Costin's all-Chinese open-weight stack, the economics of burning tokens, a fresh Salesforce OAuth breach, and jellyfish UFOs over Iran. Cast: Juan Andres Guerrero-Saade, Ryan Naraine and Costin Raiu. Timestamps: 0:00 — Introductory banter, Thinkst Canary sponsorship 2:55 — Why threat intel analysts are built for the AI moment 11:09 — Government takes the wheel: Mythos, Fable & the frontier labs 16:15 — Did the government go too far/not far enough? 25:42 — Anthropic's "best PR campaign in history" 31:52 — Alibaba, distillation & the model-router cartel 40:58 — Costin's stack: Chinese open-weight models & token economics 46:12 — Dumping, evals & the real work of AI engineering 1:04:32 — Soft power: how the world gets pushed toward China 1:14:43 — "The bullshit": over-refusal & the Opus 4.8 regression 1:32:03 — The trillion-dollar IPO endgame 1:35:49 — The Klue OAuth breach and secure-by-default 1:45:32 — Shout-outs: UAP jellyfish, LABScon 2026Links:Transcript Thinkst Canary Anthropic accuses Chinese rival Alibaba of illicitly extracting AI capabilities USG Executive Order on Promoting AI Innovation US allows Anthropic to release Mythos AI to 'trusted' US orgs US close to allowing Anthropic to restore Fable 5 model OpenAI: Previewing GPT-5.6 Sol NCSC on AI shift in cyber risk DeepSeek Alibaba Qwen NVIDIA DGX Spark StepFun Open AI Platform Xiaomi MiMo-V2.5 Z.ai - powered by GLM-5.2 Five Eyes Cyber Security Agencies Statement on AI Klue Security Incident Cybercrime Breaches Klue: Salesforce Data Impacted for Many Victims OAuth2 - OWASP Cheat Sheet Series OAuth 2.0 Policies (Google for Developers) Downed US pilot reported seeing Iranian drones swarm in ‘jellyfish’ formation LABScon 2026
Transcribe →
Katie Moussouris on the Anthropic Export-Control Mess
Jun 1901:38:24Tap to summarize(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 102: Software export controls expert Katie Moussouris joins the show to unpack the US government's abrupt move to suspend access to Anthropic's most powerful models over a so-called "jailbreak" that, on reading the paper, turned out to be a model doing exactly what defenders are supposed to do. We dig into the export-control chaos, the chemical-weapons framing of cybersecurity, the China question, and why Microsoft just resurrected a disclosure term the industry buried fifteen years ago. Cast: Katie Moussouris, Juan Andres Guerrero-Saade and Ryan Naraine. Costin is traveling. Timestamps: 0:00 - Introductory banter 1:00 - Export Controls: Fable 5 and Mythos 5 suspended 3:40 - The Anthropic–USG relationship and USG’s surveillance claim 9:40 - Self-owns, doomsday cults, and why the guardrails are "so broad" 12:42 - What the Amazon paper actually says ("fix this code") 20:33 - The chemical-weapons framing problem 23:39 - The China question and the SK Telecom angle 41:17 - Why hasn't the paper been published? 57:01 - "Free Fable": are Chinese models only months behind? 1:00:13 - The unforgiving internet and the security poverty line 1:11:18 - Microsoft brings back "responsible disclosure" (and threatens researchers) 1:29:04 - Luta Security, the AI bug flood, and shout-outsLinks:Transcript Katie Moussouris | LinkedIn JAGS on NPR: Can computer hackers get inside your mind? Anthropic Statement on the USG Export Controls Moussouris: Fable 5 Export Controls Harm US Cyber Defense Anthropic's Fable Backlash (David Sacks podcast) AI imaging company Midjourney tackles MRI scanning Microsoft Threatens Vuln Researchers Microsoft blog on CVD and POC publications TLPBLACK LABScon CFP Luta Security
Transcribe →
Mythos, Fable, and Anthropic's Big Trust Problem
Jun 1201:59:10Tap to summarize(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 101: We discuss Anthropic's Mythos 5 and Claude Fable 5 release and the bombshell that the company was silently downgrading paid users' results, sparking a heated debate over guardrails, gatekeeping, and whether elite AI reasoning is becoming a privilege for the few. Plus, AI-generated N-day exploits killing the patch window, a record-shattering Patch Tuesday, Meta's latest court filing against spyware maker NSO Group, the return of cyber paleontology, and a detour into the new government UFO drops. Cast: Juan Andres Guerrero-Saade, Ryan Naraine and Costin Raiu. Timestamps: 0:00 - Introductory banter 3:22 - The Mythos 5 / Claude Fable 5 release 14:42 - Anthropic’s silent downgrade trust problem 26:18 - Anti-competitive behavior & the AV "stealing detection" parallel 32:29 - Distillation, China & the real motive 38:04 - "Too dangerous to release" & gatekeeping vs. guardrailing 45:53 - Is Mythos a threat to malware-analysis startups? 48:20 - Dario's AI regulation essay 56:48 - N-day exploits and death of the patch window 1:07:18 - Patch Tuesday and 10x vulnerability surge 1:10:34 - Meta catches NSO Group 1:14:45 - Cyber paleontology, Shadow Brokers leaks 1:28:29 - Moonlight Maze and learning from history 1:34:22 - UFOs, UAPs and Disclosure Day Links:Transcript SemiAnalysis: Anthropic's latest model silently degrades its IQ Researchers Are Furious Over Anthropic's Hidden AI Limits Anthropic Walks Back Policy That Could Have ‘Sabotaged’ AI Researchers Anthropic: Claude Fable 5 and Claude Mythos 5 System Card: Claude Fable 5 & Claude Mythos 5 Socket: Mini Shai-Hulud, Miasma, and Hades Worms Dario Amodei: Policy on the AI Exponential Patch Tuesday (Adobe/Microsoft) WhatsApp catches NSO Group spear phishing Department of War Publishes Third Release of Unidentified Anomalous Phenomena Files Pentagon releases 3rd batch of UFO files, detailing mysterious orb sightings Orbs over a pond (UFO video) fast16 | Mystery Shadow Brokers Reference Reveals High-Precision Software Sabotage 5 Years Before Stuxnet Penquin's Moonlit Maze LABScon Call for Papers
Transcribe →
Fast16, Fanny, and Stuxnet: Cyber Paleontology Redux
Jun 502:24:29Tap to summarize(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 100: We cover AI eating reverse engineering, the death of the malware report, running local models on the DGX Spark, where Google DeepMind stands, and whether the frontier labs will stay in cybersecurity. Plus, more on Anthropic's Mythos rollout and the thinly sourced Anthropic-NSA reports, the Fast16 sabotage of physics calculations, what researchers choose not to publish, Microsoft's bad Black Hat email, and Costin's Friday UFO files. Cast: Juan Andres Guerrero-Saade, Ryan Naraine and Costin Raiu. Timestamps: 0:00 - JAGS at InfoSecurity Europe 3:40 - Sponsor: TLPBLACK 5:54 - A roadmap for security after the AI revolution 11:01 - Stripe Atlas and how easy it is to start a company 15:00 - If anyone could reverse engineer anything for $5 19:49 - Layoffs at Google's Threat Intelligence Group 21:06 - The death of reading the report 27:53 - Pitting the AI models against each other 32:07 - Grok, local models, and the DGX Spark 39:27 - Where is Google DeepMind? 45:29 - Will the frontier labs stay in cybersecurity? 52:41 - Mythos, Project Glasswing, and the NSA deal 1:16:33 - FAST16, Stuxnet, and sabotaging Iran's bomb 1:57:52 - Microsoft, Black Hat, and the chilling effect 2:14:14 - Shout-outs, UFO files, and 100 episodes Links:Transcript NSA using Anthropic's Mythos despite blacklist Anthropic: Mapping a year’s worth of AI-enabled cyber threats LLM ATT&CK Navigator (Anthropic) Ruben Santamarta on Fast16 sabotage malware A Fanny Equation: “I am your father, Stuxnet” TLPBLACK LABScon Call for Papers Disclosure Day | Official Teaser Disclosure Day
Transcribe →
Microsoft Threatens Vuln Researchers; Shadow Brokers Revisited
May 3001:59:45Tap to summarize(Presented by Ent.ai: Ent delivers intent-aware security that protects every action, adapts to every workflow, and works for every user. Enterprise threat detection, reimagined.) Three Buddy Problem - Episode 99: Microsoft is now threatening legal action against researchers who drop zero-days. We debate whether it's a fair line against extortion, or amateur-hour PR from a company that already torched its own research community? Costin plays reluctant defender, JAGS says the damage was done years ago, and Ryan reopens the long history of silent fixes and stolen bounties. Plus, on the 10th anniversary of the Shadow Brokers leak, we discuss some enduring mysteries, theories on attribution and an interesting trail that leads to Edward Snowden. We also unpack Rob Joyce's warning that China's cyber explosives are already planted in US infrastructure, and the Pope's warnings about around artificial intelligence. Cast: Juan Andres Guerrero-Saade, Ryan Naraine and Costin Raiu. Timestamps: 0:00 - Introductory banter 2:03 - The Pope's AI paper 3:35 - New sponsor: Brandon Dixon's Ent Security 9:34 - Costin's Chinese-model OSINT rabbit hole 13:34 - Codex, GPT-5.5, and the "American AI welfare state" 23:20 - Microsoft threatens vulnerability researchers 27:06 - Is it extortion or retribution? The disclosure fight 40:48 - How Microsoft's consultant class broke MSRC and MSTIC 48:42 - Silent fixes, stolen bounties, and the marketing machine 1:02:29 - Ten years of the Shadow Brokers 1:14:20 - The Snowden theory 1:32:34 - Rob Joyce: China's cyber explosives are in place 1:53:26 - Shout-outs Links:Transcript MSRC threat against security researchers Nightmare Eclipse Dennis Fisher: The Past is Always Present in Vulnerability Disclosure Introducing Claude Opus 4.8 Stolen Device Protection for iPhone Rob Joyce: China’s Cyber Explosives are in Place. Where’s our Response? China AI Security Testing Contest Brandon Dixon parses the AI security hype Ent -- Intent-Aware Security for the Enterprise Ent on LinkedIn
Transcribe →
Aaron Portnoy on Pwn2Own, the End of Easy Bugs, and AI-Fueled Offense
May 2700:40:09Tap to summarize(Presented by TLPBLACK: A cybersecurity intelligence platform focused on sharing curated, high-sensitivity threat insights and research with trusted security professionals.) Three Buddy Problem x Ekoparty Miami: Aaron Portnoy (Zero Day Initiative alum, early Pwn2Own organizer, and now at Mindgard) joins us at Ekoparty Miami to reminisce on the early days of the hacking contest, where vulnerabilities actually live (the boundaries between systems, not inside them), why LLMs will take out the trash but can't dream up the next speculative-execution-class bug, and the coming patching apocalypse when discovery 10x's overnight. Plus, why your SOC is a forensic historian, the promise of hijacking an attacker's reward loop with deception tech, and the legendary story of carrying a Walmart "fat stack" of cash to bootstrap Ekoparty in Buenos Aires. Cast: Juan Andres Guerrero-Saade, Ryan Naraine and Aaron Portnoy. Timestamps: 0:00 — Introductory banter 1:17 — Dropping out, iDefense, and getting good at reversing everything 2:19 — How Pwn2Own got started 4:15 — The most impressive Pwn2Own ever: Nils, VUPEN, and exploit "art" 5:59 — "iPhone hacked in 30 seconds" — and the 18 months behind it 6:41 — Does Pwn2Own still have a place in the AI era? 9:16 — Why LLMs take out the trash but can't invent the next bug class 12:48 — Will LLMs deliver new mitigation classes? Aaron's skeptical 18:34 — The place of the human when the easy bugs run dry 21:08 — Cognitive offloading, Halvar's warning, and skill rot 22:39 — Decompiling 800k functions: Aaron's LLM "holy shit" moment 25:26 — The patching apocalypse and why "assume breach" breaks 28:15 — Compounding asymmetries: why offense just transcended defenseLinks:Transcript Aaron Portnoy | LinkedIn Mindgard - Automated AI Red Teaming Pwn2Own Ekoparty Miami Nils2Own: 'I want to see security flaws fixed' VUPEN Charlie Miller on hacking iPhones, Macbooks, Cars LABScon 2026 TLPBLACK
Transcribe →Perri Adams on Proof Engines, LLMs, and the New Era of Verifiable Code
May 2600:40:27Tap to summarize(Presented by TLPBLACK: A cybersecurity intelligence platform focused on sharing curated, high-sensitivity threat insights and research with trusted security professionals.) Three Buddy Problem x Ekoparty Miami: Perri Adams of DARPA AIxCC fame joins the show to chat about proof engines, formal methods, and why LLMs just made a once-niche corner of computer science suddenly essential. We get into why verifiers and proof engines are the key to effective AI, why vulnerability research is so far ahead of threat intel, and the case for baking security checks directly into code generation tools like Claude Code and Codex. Plus, designing a multi-million dollar challenge that's allowed to fail, the Mythos "too dangerous to release" debate, and musings on every LLM-discovered bug being a public bug by default. Cast: Juan Andres Guerrero-Saade, Ryan Naraine and Perri Adams. Timestamps: 0:00 — Introductory banter 1:09 — Why LLMs just made formal methods relevant again 4:03 — Proof engines, explained 8:43 — Can a layman grab this fire? The calculus problem 11:58 — Vuln researchers are scrappy kids with a trust fund 14:55 — Pitching AIxCC inside DARPA: hard sell or easy sell? 18:00 — Designing a challenge that's allowed to fail 22:06 — Inside Team Atlanta's 150-page winning system 24:00 — Why this is bigger for defense than for offense 31:49 — Mythos, safeguards, and "every LLM bug is a public bug"Links:Transcript Perri Adams (@perribus) / X DARPA AIxCC - AI Cyber Challenge AIxCC Final Competition Winners Announcement Team Atlanta (AIxCC) Team Atlanta AIxCC Research Publications and Source Code DARPA's Perri Adams on CTF hacking, new $20M AI Cyber Challenge Microsoft MDASH OffensiveCon25 Keynote (Perri Adams) Binary Analysis: An AI Success Story TLPBLACK LABScon 2026
Transcribe →