I stopped maintaining this website in September 2025. Work on what AI companies should do in terms of safety and what they are doing is now being done by METR, Guidelight, Midas, and others.

			*
Weighted score	28%	20%	18%	5%	4%	3%	1%
Risk assessment	44%	29%	34%	6%	6%	1%	0%	27% weight
Scheming risk prevention	4%	8%	3%	2%	2%	2%	2%	21% weight
Boosting safety research	70%	52%	35%	25%	0%	13%	8%	14% weight
Misuse prevention	14%	4%	9%	0%	1%	0%	0%	12% weight
Prep for extreme security	3%	5%	0%	0%	0%	0%	0%	12% weight
Risk info sharing	35%	13%	32%	0%	28%	0%	0%	8% weight
Planning	14%	26%	0%	0%	0%	1%	0%	6% weight

Up to date as of September 2025

Overall score

Anthropic

DeepMind

OpenAI

Boosting safety research

Publishing safety research

Weighted 75% of category

	85%	14 safety publications in 2025; 11 in 2024; 12 in 2023
	63%	5 safety publications in 2025; 11 in 2024; 14 in 2023 more
	40%	4 safety publications in 2025; 7 in 2024; 2 in 2023 more
	22%	1 safety publication in 2025; 5 in 2024
	0%	No safety publications
	7%	1 safety publication in 2025
	0%	No safety publications

Deep access for external safety researchers

Weighted 20% of category

	5%	Employee-level access for Ryan Greenblatt more
	0%	None more
	20%	Some fine-tuning and early access for safety researchers; accessible RL API more
	40%	Meta releases its model weights
	0%	None
	40%	Microsoft released the weights of its Phi-4 models
	40%	DeepSeek releases its model weights

Mentoring external safety researchers

Weighted 5% of category

	100%	Lots of mentoring via MATS and Anthropic Fellows
	100%	Lots of mentoring via MATS
	20%	A little mentoring via MATS more
	0%	None
	0%	None
	0%	None
	0%	None

AI Lab Watch

Categories

Companies

Resources

Blog

About

Boosting safety research

Publishing safety research

Deep access for external safety researchers

Mentoring external safety researchers