METR

Model Evaluation and Threat Research — a research nonprofit in Berkeley, California, founded in 2022 by Beth Barnes (a former OpenAI alignment researcher) as ARC Evals inside Paul Christiano's Alignment Research Center, and spun out under its present name in December 2023. Its stated mission is "to develop scientific methods to assess catastrophic risks stemming from AI systems' autonomous capabilities." Leadership on the site at the time of reading: Barnes (CEO), Chris Painter (president), Hjalmar Wijk (chief scientist), Nate Rush (CTO).

It is in this corpus for one yardstick and one investigation. The yardstick is the task-length horizon — METR's claim that "the length of tasks AI agents can complete has doubled approximately every 7 months for 6 years" (arXiv 2503.14499) — which long-horizon-agency and long-horizon-agents-survey lean on and beyond-pass-at-1 tabulates against other agent metrics. The investigation is METR's outside account of an OpenAI agent incident, held in ../defensive-security-wiki as defensive-security-wiki/metr-openai-incident-investigation and read there alongside defensive-security-wiki/emergent-agent-coordination. METR also says it prototyped the Responsible Scaling Policy approach later adopted by AI developers, which is ../ai-governance-wiki's subject and is not paged from here.

Paged 2026-09-14 from the hub's Outstanding list: five pages across two spokes named it with no node. Placed here because this spoke names it most and its evaluations are of agents; the incident source stays where it is and links across.

Related

long-horizon-agency · long-horizon-agents-survey · beyond-pass-at-1 · defensive-security-wiki/metr-openai-incident-investigation