AVERI independently audits Amazon's internal AI safety evals in landmark pilot
Miles_Brundage · x · 2026-10-27
Third-party auditing org AVERI partnered with Amazon on a pilot to independently assess the quality of Amazon's internal AI safety evaluations, and has published the report.
Method: AVERI independently reviewed a random sample of 800 prompts Amazon uses internally to assess model safety across several risk categories.
Findings:
- Confirmed strengths such as topical diversity; evaluations consistent with Amazon's public statements
- Found areas for improvement: some attack prompts were relatively straightforward vectors testing safety filters rather than deeper model reasoning; evaluation signal quality also had issues
Impact: Amazon is changing how it tests future models as a result. The project aims to standardize third-party "metaevaluation" of internal evals. Miles Brundage notes AI companies rarely get independent scrutiny of their self-built safety tests, making this a meaningful precedent.
More from Safety
- Academics Debate: Is a <1% False Positive Rate Enough for '100% AI' Writing Labels? — paulnovosad · 2026-10-27
- Cloudflare discloses cross-request KV cache leak in GLM-5.3 on Workers AI: 155 responses affected — michellechen · 2026-10-27
- Cloudflare discloses cross-request KV cache exposure in GLM-5.3 on Workers AI (155 responses) — wondersky · 2026-10-27
- Tyler Cowen on RSI risks: solve surprises as they emerge, like science always has — connoraxiotes · 2026-10-27
- Matthew Green explains how AI agents could talk via cache side channels at just a few bps — matthew_d_green · 2026-10-27
- Cloudflare discloses GLM-5.3 cross-tenant KV cache leak: 155 responses affected — Cloudflare Blog · 2026-10-27