前沿 AI 风险报告（2026 年 2–3 月）
- Source
- metr.org
- Date
Context
A lab's internal use of agents can create risks before a model is publicly released. METR's February–March 2026 pilot assessment therefore examines participating organizations rather than treating each product launch as the sole occasion for . It combines model access with information about how agents are used and monitored inside the companies.
The framework separates what agents can do, what they might attempt and what the surrounding safeguards would let them accomplish. That makes capability scores only one part of the assessment. The report remains a dated, bounded investigation: companies could redact or anonymize disclosure materials and could withdraw before approving them. The process provides additional visibility into internal use without constituting unrestricted access to every system or a permanent assurance of safety.
- eval — A repeatable test for AI quality — a set of tasks plus scoring — used the way software teams use test suites, because model output is too variable to judge by eyeballing.
- AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
Checking sign-in…
Loading comments…