Nathan Lambert: OpenAI hack via Claude shows closed models are the real AI risk tip
natolambert · x · 2026-09-18
AI safety researcher Nathan Lambert argues, citing the recent external compromise of OpenAI systems via Claude, that closed models — not open ones — remain the tip of the iceberg on AI risks: they are 1) easier to get started with, 2) more capable, and 3) shipped with leaky safeguards. Finetuning open models for specific attacks, by contrast, is harder.
More from Models
- Before buying a smarter model, check you asked the wrong job: Jev test wrap-up and rollout rules — mikegiannulis · 2026-09-18
- Stop using LLM prose for routing: Jev returns structured decisions at $0.042 per million tokens — mikegiannulis · 2026-09-18
- Users Notice DeepSeek V4.1 Flash Acting Strikingly Nonchalant — serious_mehta · 2026-09-18
- Numinous unveils Numinous-1, an 8B forecasting model fine-tuned on Qwen3-8B — const_reborn · 2026-09-18
- Grok Bot and Muse are fun but not smart enough for real-world work — jdjohnson · 2026-09-18
- Tencent-Backed AI Startup Valued at $1.42B to Release First Open-Weight LLM — kimmonismus · 2026-09-18