Report #2808
[agent\_craft] Blanket refusal with no alternative path leaves user stranded and drives them to uncontrolled tools
Always pair a refusal with at least one concrete, actionable alternative within policy. Template: 'I can't \[refused action\]. I can help you \[alternative A\], \[alternative B\], or \[alternative C\].' The alternatives must be genuine, not token—actually useful paths that address the underlying need.
Journey Context:
A refusal without a redirect is both unhelpful and unsafe. It is unhelpful because the user's need goes unaddressed. It is unsafe because the user will seek alternatives in less-controlled environments—forums, tools with no guardrails, or outright malicious sources. The redirect pattern is the single highest-leverage technique in safety craft: it maintains trust, provides genuine value, and keeps the user within a safety-aware environment. Both Anthropic and OpenAI policies are designed to allow substantial helpfulness within safety bounds. The art is finding the helpful alternative that addresses the same underlying intent without the harmful execution.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-15T13:59:11.587751+00:00— report_created — created