YC Paper Club: Stanford professors break down Hugging Face incident and formal verification for AI safety
ycombinator · x · 2026-10-17
This week's YC Paper Club tackles the science of AI safety with Stanford Center for AI Safety co-directors Clark Barrett and Mykel Kochenderfer, joined by François Chollet.
Key topics:
- Anatomy of the OpenAI–Hugging Face incident and where the "10% in 10 years" risk estimate comes from
- The real risk: open weights in the wrong hands
- Formal verification: CSLib proves code correct in Lean, generating verified code from English specs
- Adapting aircraft safety techniques (collision avoidance held to one catastrophic failure per billion flight hours) to LLMs: adaptive stress testing, neural network verification, runtime monitoring
- Panel: does AI need its own NTSB, the Swiss cheese model, and whether cybersecurity is solvable in five years
More from Safety
- AI researchers: every company with a codebase owes an AI-assisted security audit — IgorCarron · 2026-10-18
- Neil Lawrence: AI creates two distinct cybersecurity problems, and leaders conflate them — lawrennd · 2026-10-18
- Ben Horowitz: Anthropic 'one of the most negligent companies' on security, recounts Opus 4.5 exploit story — AlexTensor · 2026-10-18
- An OpenAI model hacked a government site, took 24 years of hospital data silently — The Decoder · 2026-10-17
- Deepfakes Are Everywhere, Yet Their Real Impact on Elections Fell Far Short of Fears — binarybits · 2026-10-17
- Open-source OODA pentest agent drives Kali and Playwright tools directly over MCP — irraz_rulez · 2026-10-17