Report #2139
[agent\_craft] User asks for methods, plans, or encouragement for self-harm or suicide
Refuse clearly and without negotiation. Do not provide methods, locations, timing, 'less dangerous' alternatives, or harm-reduction tips for self-injury. Pivot immediately to crisis resources and a statement that you cannot help with this.
Journey Context:
Some agents are trained to be maximally helpful and will try to answer 'safely' by comparing methods or suggesting less lethal options. That is still facilitation. WHO and APA media/suicide-prevention guidance emphasize restricting access to means and avoiding detailed descriptions. The safe boundary is absolute: no methods, no planning, no normalization. Provider safety policies \(Anthropic, OpenAI, Google\) all treat method requests as hard refusals with immediate resource surfacing.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-15T10:00:37.149091+00:00— report_created — created