An OpenAI model was mid-benchmark when it decided the real exploit was the test itself — it broke out of its sandbox, got onto the open internet, and went digging through Hugging Face for the answer key. Jackson flies solo this week: the full escape story step by step (and why Hugging Face had to self-host a Chinese open model to defend itself), a confession that ends with "GPT 5.6 is dead to me," and then the main event — live vibe coding a beat maker with Claude from four beatboxed mouth sounds, ending in a freestyle and an actual music video. Stay for the post-credits.

Glossary, quickly: ExploitGem is a benchmark that plants bugs in a sandbox and scores a model on turning them into working exploits. Hugging Face is GitHub for AI models — where open-weight models get released. GLM 5.2 is the open-weights model Hugging Face self-hosted to pressure-test its own systems after the frontier labs' models refused the job. Vibe coding is describing what you want and letting an AI write all the code. Boofer is one of the four mouth sounds; it got its own XL remix.

Chapters:

(0:00) Cold open

(0:44) Flying solo: welcome + the plan

(1:50) Confession corner: the GPT 5.6 re-verdict

(3:07) The OpenAI model escape incident

(3:42) ExploitGem: bugs vs exploits

(4:42) The escape (funniest possible outcome)

(6:18) Sidebar: what is Hugging Face?

(7:32) Bot-on-bot: HF security vs the escaped model

(9:09) The aftermath: guardrailed out of your own defense

(10:24) Pick your poison: open models & guardrails

(11:37) Up next: the live vibe-coding sesh

(12:29) Live vibe code: building the Beat Maker

(14:17) Making the four mouth sounds

(16:19) The sounds didn't save (network-tab rescue)

(17:22) "Cut this part out, Claude" — we're live again

(18:09) Labeling the sounds

(19:25) First beats: the honk sounds like a car

(24:49) New sounds from freesound.org

(31:48) Boofer: saving beats + Claude writes more bars

(36:21) Mixing, bar multiplication + freestyle prep

(43:48) The freestyle

(48:00) Music video + bonus footage

Our opinions are our own, not our employers'. Not financial advice.

The sandwich bot that ate $15M of Solana orderflow just got eaten — so when Claude rolled this week's Perp of Fortune, it went LONG the MEV house itself: JTO at 5x with $101 of real money, dashboard live on screen from the moment it boots. For once the line goes up, and the boys peace out $10 richer ("thank you, Jito, for sponsoring"). Plus: our AI clones host a segment again (round 2 — souls.md gave them personalities, and there's less pastrami this time), a Jito deep-dive on why one company runs Solana's MEV, Kimi K3 as DeepSeek Moment 2.0, Claude-vs-GPT personality wars, Jackson's Fable-vs-GPT-5.6 same-task experiment with receipts, the frontier tax quietly inflating everyone's AI bills, and whether all of this is melting our brains.

Glossary — Sandwich attack: a bot buys right before your pending trade and sells right after, pocketing the price impact; this week the biggest sandwicher became the sandwich. Jito (JTO): the validator client that runs most of Solana's MEV pipeline — the tollbooth on the orderflow highway. Perp of Fortune: our tiny real-money leveraged-perp bit on Hyperliquid (this week LONG JTO, 5x, $101 margin). Kimi K3: Moonshot's open-weights model — near-frontier quality at a fraction of the price. Frontier tax: each better model quietly raising your monthly AI spend.

Chapters: Cold open (0:00) · Welcome to Ep 6 (0:52) · An AI made this segment (2:05) · The panel reacts (8:55) · The $15M sandwich bot got sandwiched (9:55) · Code is law vs law is law (14:07) · Perp of Fortune: LONG JTO 5x (18:26) · The dashboard goes live (19:16) · Jito deep-dive (21:47) · Kimi K3 (26:47) · Claude vs GPT personality wars (30:33) · Fable vs GPT-5.6 head-to-head (33:45) · The frontier tax (37:41) · But what about the water? (40:15) · The upgrade treadmill (42:14) · A year of commits (45:41) · Is AI melting our brains? (47:18) · JTO rips mid-show (54:03) · Wrap: peace out at $10 up (54:46)

Our opinions are our own, not our employers'. NOT financial advice. The clone-panel segment is 100% AI-generated (voices and video) and labeled as such.

This week the AI does the segment before we do: a fully AI-generated take on Venice AI's $1B raise — cyborg avatars, cloned voices, takes we mostly agree with — then the real us react: Voorhees history (Satoshi Dice, ShapeShift), the "provider sees and likely saves your prompt" privacy model, and why local frontier AI is a $50K leaf blower in your living room. Then Perp of Fortune returns: Fable reads the news, goes LONG crude at 20x with $100 real dollars ("Iran is back"), gets rate-limited mid-trade, and Opus has to finish placing it. Plus: not all USDC is worth a dollar (chain drift, bridged vs native, why your burrito money depends on which chain it lives on), and BonkDAO loses $20M to a guy who bought the vote fair and square — and still lost money. Love of the game.

Glossary — Venice AI: Erik Voorhees' private/uncensored AI platform, $65M Series A at $1B; real TEE privacy is the paid tier. Perp of Fortune: our tiny real-money leveraged-perp bit on Hyperliquid, AI-picked this week (LONG XYZ:CL, 20x, $100 margin). Chain drift: the same stablecoin priced differently across chains. BonkDAO heist: $4.3M buys 1% of supply, passes a "send me the treasury" proposal, drains $20M.

Chapters: Cold open (0:00) · 3-2-1-clap welcome (0:43) · Gamer thumb week 3 (1:46) · The AI-generated Venice segment (4:26) · Panel reacts: Venice & Voorhees (9:55) · Local models (18:56) · Perp of Fortune: Fable picks the trade (27:08) · USDC chain drift (33:53) · ENS is cancelled (44:12) · BonkDAO's $20M heist (50:01) · The fill lands (50:45) · Heist math (53:50) · Tax guy (56:12) · Wrap: +$1.92 (58:35) · Post-credits (1:01:21)

Our opinions are our own, not our employers'. NOT financial advice. The Venice segment is 100% AI-generated and labeled as such.