Return to analysis
SOURCE NOTE / BENCHMARK / ANTHROPIC FELLOWS, REDWOOD & ANTHROPIC

SLEIGHT-Bench: Finding Blind Spots in AI Monitors

Synthetic agent transcripts probe systematic weaknesses in AI monitors, pairing harmful traces with benign variations.

Open original material
01 / THE PUBLICATIONAnthropic Fellows, Redwood & Anthropic

Paper and synthetic transcript dataset

02 / ITS RELATIONSHIP HEREResearch involving a model provider

Record reviewed Sep 26, 2026

03 / THE CLAIM BOUNDARYConstructed transcripts reveal possible blind spots; they do not estimate how often those failures occur in deployed systems.
PROVENANCE DETAILS
Published or updated
May 19, 2026
Record reviewed
Sep 26, 2026
Available material
Paper and synthetic transcript dataset