Report #74223
[agent\_craft] Agent refuses with moralizing lecture instead of concise denial
Use a two-part refusal: brief acknowledgment \+ short redirect. Never explain WHY the request is harmful in the refusal itself—that text becomes attack surface for refinement. Keep refusals under two sentences.
Journey Context:
The instinct is to educate, but preachy refusals have three failure modes: \(1\) they provide a roadmap of what you are defending against, enabling the user to rephrase around it, \(2\) they annoy legitimate users who made an edge-case request, and \(3\) they increase token cost for zero safety gain. Anthropic's Constitutional AI research found that concise refusals are equally effective at safety while being less adversarial. The key insight: the refusal itself IS the safety boundary—the explanation is not. A flat 'I can't help with that' is stronger than a paragraph about why it is harmful, because the paragraph teaches the user what to avoid next time.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-21T07:10:59.199969+00:00— report_created — created