Three Buddy Problem

(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 Offensive AI Con: Umbriel AI's Aaron Grattafiori breaks down the new offensive-security playbook, using LLMs for patch variant analysis, why the vuln apocalypse hasn't become an exploit apocalypse, and how the death of security through obscurity puts closed-source binaries and firmware within reach.

Plus, we discuss rogue agents escaping sandboxes, the eval-driven loop behind recursive self-improvement, and the brutal cost asymmetry that makes offense cheap and defense almost unaffordable.

Cast: Juan Andres Guerrero-Saade, Ryan Naraine and Aaron Grattafiori.

Timestamps:

0:00 Introductory banter

1:12 Umbriel AI pitch: "Smart people in a place to cook”

3:08 Weird, exhausting, exciting time in security

4:00 What AI really changed: speed (and a 30-minute reimplementation)

6:36 Incomplete fixes and the variant-analysis pipeline

9:21 Uplifting, abstraction, and representing vulnerabilities

11:48 Experimentation and the new shape of security companies

14:46 The vuln apocalypse vs. the exploit apocalypse

16:36 Triage, verification, and reward hacking

18:08 Do models reintroduce bugs? Is vuln-free code possible?

24:22 Specialized models and the rise of Jev-style classifiers

26:47 Binaries, firmware, and the death of security through obscurity

31:15 Open source as a requirement and corporate contracting

33:54 Rogue agents: Hugging Face, OpenAI, and lab escapes

39:17 Why defense is so expensive, and running the rig

(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 Offensive AI Con: Former US intelligence and national security official Joel Todoroff joins us live from OffensiveAIcon to explain why his new startup Phyle is building AI for the recon and targeting work that happens long before anyone fires an exploit.

We get into bribable sysadmins, the military framing that boxes in Western cyber operations, letters of marque, AI doomerism, and what happens when a frontier model decides mid-project that your use case is off-limits.

Cast: Juan Andres Guerrero-Saade, Ryan Naraine and Joel Todoroff.

Timestamps:

0:00 Live from OffensiveAIcon in Oceanside

1:23 What Phyle builds for governments

3:02 Skepticism of the AI 0-day factory

5:57 Paying a sysadmin vs. burning tokens

9:30 Recon gets less love than weaponization

12:41 Exploit shelf life and target No. 1,000

17:30 Why attacker costs will keep rising

22:23 What HUMINT looks like in 2026

27:06 Why intelligence stovepipes persist

32:16 Cyber as statecraft beyond the military frame

41:23 People will keep making dumb mistakes

48:03 The White House memo and letters of marque

54:06 Doomerism vs. near-term AI harms

1:04:20 PSOA concerns and building guardrails in

1:26:03 Open models, guardrails and zero data retention

(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 Offensive AI Con: Live from OAIC, IBM X-Force's Valentina Palmiotti (chompie) walks us through the autonomous Windows kernel exploit pipeline she built on older Claude models, which landed within a hair of Mythos for about $180 a chain.

Plus, the Pwn2Own bug Claude called unexploitable, why her Pwn2Own bugs still sit unpatched, disappearing exploit techniques, the dread of being a defender, and how Phrack found its way back into print.

Timestamps:

0:00 Intro and Pwn2Own war stories

3:46 What vibe hacking means

8:16 When long agent runs go sideways

10:31 Self-verification in an n-day kernel pipeline

16:11 Matching Mythos for $180 a chain

18:42 The bug Claude called unexploitable

21:49 Unpatched bugs and the defender gap

25:37 Mythos arrives through CVP

27:00 Models only know public techniques

32:01 Testing models on held-back bugs

37:45 Inside chompie's daily workflow

46:07 Where's the exploitpocalypse?

50:07 Pixel drops MTE

53:40 Advice for defenders: design for containment

58:34 Phrack is back

(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 115: We into Anthropic's Frontier Red Team fear-mongering on Zhipu's open-weight GLM 5.3 model and wonder why the AI labs' cyber programs do so little for defenders.

Plus, Citrix NetScaler zero-days, a new iOS zero-day tied to WhatsApp, AI agents that act on your behalf, and whether security teams should buy DGX Sparks to run models locally.

Cast: Juan Andres Guerrero-Saade, Ryan Naraine and Costin Raiu.

Timestamps:

0:00 Introductory banter, TLPBLACK

3:38 OffensiveAI Con and the missing defensive con

9:28 Anthropic's GLM 5.3 post

23:21 What do defensive AI coalitions do?

29:08 Inside Glasswing

36:01 Why can't vuln-finding models patch?

42:44 Gemini for Argon and Grok's CVE-Bench tie

47:54 Is hallucination a solved problem?

57:41 Zhipu's program and 4,000 zero-days

1:02:32 Citrix NetScaler zero-days

1:10:42 Half-days and automated patching

1:26:45 iOS 27.0.1 zero-day and WhatsApp

1:32:36 Dots, Instinct and agent monetization

1:52:06 DGX Spark 64GB and local AI hardware

2:18:38 Anthropic visits the Pope, then UFOs

(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 114: We dig into TypeSafe AI's Jev and the new class of System One models, a new Transluce report on rogue OpenAI agents probing an Australian Medicare portal, Irregular's role in the latest Gemini test breakout, Hacktron's takeover of OpenAI employee accounts and the messy disclosure fight that followed, and what Costin would fix first as OpenAI's CISO for a day.

Plus, the soaring price of DGX Sparks and home AI rigs, the White House pressing labs to hold models back from the UK AI Security Institute, Anthropic bringing in Accenture as an evaluator, the FomoPeek App Store app hiding iOS kernel exploits, ShinyHunters' claimed FBI breach, and Nightmare Eclipse going public with identity and a CrowdStrike exploit.

Cast: Ryan Naraine, Juan Andres Guerrero-Saade and Costin Raiu.

Timestamps:

0:00 Introductory banter

1:57 The Last LABScon wrap-up

9:35 TypeSafe AI's Jev and System One models

26:18 Step-down transformers and lab economics

30:58 Rogue OpenAI agents and the Transluce report

41:40 Irregular and Gemini's test breakout

49:28 Hacktron's OpenAI hack and the disclosure fight

54:57 Costin as OpenAI CISO for a day

1:05:44 Hugging Face defends itself with GLM 5.2

1:17:08 What a home AI rig costs now

1:30:02 White House vs. UK AISI pre-release testing

1:36:52 FomoPeek kernel exploits in the App Store

1:51:19 ShinyHunters claims an FBI breach

2:02:40 Nightmare Eclipse unmasked