Apollo Research wants AI safety claims that outsiders can test
Two new proposals shift attention from a clean test result to access, evidence and the ability to find a counterexample.
2026-10-05 · 4 min · 2 sources
Apollo Research
x · Follow ↗
Evals & red teamingAlignment & interpretability
Scheming, deception, model evaluations and control
- Type
- Organization account
- Priority
- Tier 1 · core
- Perspective
- Independent evaluations
- Best for
- Research releases and technical discussion
- Publishes
- Frequent
- Editor's note
- Current official handle; follow linked reports.
Our analysis featuring Apollo Research
Recent from Apollo Research
- We're excited about embedded evaluations. Any developer should be able to show its models are not sc 2026-10-01
- Access limited to a scope agreed in advance can't account for unforeseen risks, like a new training 2026-10-01
- Assessing all four requires persistent, deep access: training techniques (including experimental one 2026-10-01
- Claim 4 is about monitorability. Evaluating the first three claims hinges on understanding why a mod 2026-10-01
- Claim 3 is about internal deployment. A lack of detected incidents only means something if monitors 2026-10-01
- Claim 2 is about evaluations. They need coverage, volume and realism, should run throughout training 2026-10-01