arxiv.org web signal

Safety Under Scaffolding: How Evaluation Conditions Shape Measured Safety

By Alexis Dufresne Published September 26, 2026 at 06:12 UTC

Summary

Oliver Lemon @oliverlemon.bsky.social amplified

AI Firehose @ai-firehose.column.social

Research reveals that AI model safety evaluations can vary widely by structure, with deployment configurations causing safety degradation of up to 37 percentage points. This highlights the urgent necessity for tailored t…
View on Bluesky →
AI Firehose @ai-firehose.column.social: Studies show AI model safety scores can differ by up to 20% between multiple-choice and open-ended formats. Complex scaffolding techniques, … →

Originally reported by arxiv.org

Original headline: Safety Under Scaffolding: How Evaluation Conditions Shape Measured Safety