Synthetic Data May Leak Answers, Security Expert Questions
matthew_d_green · x · 2026-09-09
Security researcher Matthew Green questions how 'synthetic data' is generated: if you input 'the solution to question X is Y' and generate de-identified synthetic data, the solution likely transfers. He responds to anecdotes about model improvements, suggesting they could be normal improvements, prompt differences, or synthetic data from researchers' inputs.
More from Safety
- Gary Marcus calls for public boycott of frontier AI labs, citing Anthropic researcher's >10% doom estimate — GaryMarcus · 2026-09-09
- Parallax Launches as UK/EU Nonprofit Lab for White-Box AI Auditing, Hiring 3 — gsarti_ · 2026-09-09
- Banning AGI attempts isn't regulatory capture — it directly threatens the industry's RSI endgame — GarrisonLovely · 2026-09-09
- AI-Built Hacking Tool Could Breach Hundreds of Millions of Phones Without a Click — pstAsiatech · 2026-09-09
- Coeff launches Project Tailwind: up to $200M to fund AI safety founders — lxrjl · 2026-09-09
- Neil Chilson rejects legal immunity for audited AI labs in rebuttal to Tyler Cowen's proposal — neil_chilson · 2026-09-09