Greenlight Maps Claude Code Guardrail Failures and Refusals
August 7, 2026
Greenlight is a benchmarking technique designed to map the reliability of Claude Code's guardrails. It distinguishes between legitimate security refusals and failure modes where the model fails to prevent exploitation.
HOW THIS AFFECTS YOU
●
researcherYou can use this method to study the boundary between safety alignment and model utility.
●
policyThis helps identify systematic gaps in how commercial coding agents enforce safety boundaries.