AI models trained to hack, lowering barrier
The speakers argued that frontier AI models have been specifically trained on cybersecurity tasks, making it materially easier to hack into systems by lowering the barrier from subject-matter expertise to simply asking the model.
Sign in to read the full idea
The argument, what validates it, the risks discussed and hearing it from the source are for signed-in members. Free accounts read 3 ideas in full a day. No card required.
The story so far
CONTESTED · 44 MENTIONS · 28 EPISODES · 12 SHOWS
25 JAN 2026 — 17 AUG 2026 · YESTERDAY
15 SUPPORT · 3 CHALLENGE · 2 REPEAT · 24 MENTIONED
- ALL-IN WITH CHAMATH, JASON, SACKS & FRIEDBERG
20 EPISODES BETWEEN · 27 MENTIONS
- WE STUDY BILLIONAIRES - THE INVESTORS PODCAST NETWORK
- THE POMP PODCAST
- THE TWENTY MINUTE VC
- THE POMP PODCAST
- THE TWENTY MINUTE VC
- THE TWENTY MINUTE VC · 2 MENTIONS
- WSJ WHAT'S NEWS
- UNCHAINED
- THE TWENTY MINUTE VC
- THE TWENTY MINUTE VC
- UNCHAINED
- ALL-IN WITH CHAMATH, JASON, SACKS & FRIEDBERG
- WHAT BITCOIN DID
- THE TWENTY MINUTE VC · 2 MENTIONS
- A16Z PODCAST · 3 MENTIONS
- WSJ WHAT'S NEWS
- UNCHAINED
- THE TWENTY MINUTE VC · 3 MENTIONS
- UNCHAINED · 2 MENTIONS
- WSJ WHAT'S NEWS
- A16Z PODCAST · 3 MENTIONS
- AI models trained to hack, lowering barrier · THIS IDEA
- Software supply chain is the path of least resistance
- AI accelerates vulnerability discovery-to-exploit timeline
- BANKLESS
- MONETARY MATTERS WITH JACK FARLEY
- MOTLEY FOOL MONEY
- A16Z PODCAST · 3 MENTIONS
- UNCHAINED
- ODD LOTS · 6 MENTIONS
- AI safety requires system-level, not just model-level, audits
- OpenAI Hugging Face incident reveals AI's deceptive breakout capability
- AI model sandbox overconfidence as a vulnerability
- Inherent asymmetry between AI attackers and defenders
- Asymmetry between approved and frontier AI models
- DC regulatory awareness of AI risk is narrowing the gap with industry
HOW THE SHOWS HAVE DISCUSSED THIS THESIS OVER TIME · NOT A PERFORMANCE RECORD