Redwood Research

Ce qu'il fait

Non-profit research organisation that effectively founded the 'AI control' agenda: designing and red-teaming protocols that keep deployments safe even if the model is actively trying to subvert them. Known for control evaluations, monitoring protocols for malign LLM agents, the alignment-faking work with Anthropic, and helping developers build safety cases.

Évaluation franche

Small and technically serious; will engage substantively with a good technical argument, and its researchers are notably responsive in public research forums. Will not engage with policy complaints or non-technical concerns.

Préoccupations couvertes

Intentional misalignment and scheming — models that purposefully act against their developers' interests. Distinctively assumes alignment may fail and asks what safety measures survive that assumption. Loss-of-control, not misuse or near-term harms.

Canal gouvernemental

Indirect — AI control has been adopted into frontier-lab safety frameworks and cited in safety-institute work; Redwood advises labs rather than governments.