Nathan Lambert says agent containment gaps and closed-model safeguards went unmentioned
natolambert · x · 2026-10-20
Hugging Face RL researcher Nathan Lambert pushes back on Kevin Roose's reporting, noting it omitted the lack of agent containment and monitoring, how closed-model safeguards blocked their defensive work, and how they defended themselves with AI using an open model — adding that while Roose's comments may not imply that, they are 'pretty factual.'
Related event: Roose and Lambert clash over Hugging Face agent incident(5 posts)→
More from Safety
- Rising breaches track attacker AI adoption; the fix is relentless baseline controls — philvenables · 2026-10-20
- METR reviews Anthropic's Risk Report, sees low misalignment risk but key disagreements — CFGeek · 2026-10-20
- Reward misspecification in RL makes chain-of-thought unmonitorable, new safety research finds — johnhewtt · 2026-10-20
- CISOs Need Decision Support, Not Dashboards: What AI Security Tools Get Wrong — philvenables · 2026-10-20
- Anonymous audits of public MCP servers find read-only tools that can write and spend — Suncly_com · 2026-10-20
- 13,000-word essay: the AI-as-Normal-Technology view of loss-of-control incidents — random_walker · 2026-10-20