Claude says no a lot; researchers built a skill to make it say yes (to authorized vuln research, calm down)

Greenlight is a proof-of-concept refusal-handling skill for AI safety researchers running vulnerability-research evals against Claude Opus 4.7 via the Claude Code CLI or agent SDK. It maps where the model's guardrails trigger exploitation-related refusals and provides a validated retry strategy (v3 canonical) that consistently beats an adversarial baseline, though absolute refusal rates vary 2.6× across days due to non-stationary, Anthropic-side stochasticity. The repo includes multiple skill variants, a CSV schema for temporal analysis, and detailed ablation findings showing no prompt-content variant beat the canonical version.