Vigilance

We operate on the assumption that capability jumps are unpredictable. Continuous monitoring of every major release is mandatory, not optional.

Technical Rigor

Our evaluations are reproducible. We provide the exact prompt engineering, scaffolding, and environment configs used to elicit capabilities.

Precautionary Principle

When a model nears a critical threshold, the burden of proof for safety lies with the developer. We alert the public before the line is crossed.

Why We Exist

The development of frontier AI systems proceeds at an unprecedented pace. Labs release increasingly capable models, often with limited public disclosure of their true capabilities. Safety evaluations, when they exist, are frequently conducted internally without independent verification.

AI Safety Oversight was founded to fill this gap. We provide an independent, transparent mechanism for tracking dangerous capabilities across all major AI systems. Our work ensures that concerning developments don't slip through unnoticed.

We believe that public awareness is a prerequisite for informed governance. By surfacing and verifying capability concerns quickly, we enable researchers, policymakers, and the public to make informed decisions about AI development trajectories.

2024
Founded
42
Models Tracked
156
Incidents Verified

How We Work

1

Report

Researchers submit capability concerns through our intake system

2

Corroborate

Automated research runs within 4 hours to gather supporting evidence

3

Review

Verified members assess and validate the findings

4

Publish

Confirmed incidents are added to our public tracker

Our Network

We collaborate with leading AI safety researchers, academic institutions, and policy organizations worldwide.

47
Verified Reviewers
12
Partner Institutions
24
Evaluation Scripts

Join the Oversight Network

Whether you're an ML researcher, security expert, or concerned citizen, there's a role for you in ensuring AI safety.

Apply as Evaluator Report a Concern