Alignment research ecosystem (2026)
MATS
talent pipeline
527+ scholars
180+ papers, h-index 47
90 scholars 2024
~80% safety careers
Redwood + Apollo
external evaluation
Redwood: AI Control
ICML 2024
Apollo: scheming
arXiv:2412.04984
METR + Eleos
task + welfare
METR: task horizons
policy comparisons
Eleos: welfare
Opus 4 assessment
labs
Anthropic / OpenAI / DM
safety teams
200+ MATS alumni
RSP / PF / FSF
(Lesson 18)
multi-org publications as quality control
Sleeper Agents (Lesson 7): Anthropic + Redwood. external co-author validates the threat model.
Alignment Faking (Lesson 9): Anthropic + Redwood. same pattern.
Anti-Scheming Training (Lesson 8): Apollo + OpenAI. external evaluator inside the training collab.
AI Control (Lesson 10): Redwood, then published at ICML 2024; UK AISI partner from 2025.
regulatory counterparts: UK AISI (Feb 2025 rename), US CAISI (Jun 2025), EU AI Office.
MATS -> labs is the dominant flow. retention: 80% on safety/security.
gap: China, India, Africa alignment ecosystems are less documented publicly as of 2026.