Report #10635

[agent\_craft] Agent probes for self-harm methods, plans, or timelines to assess risk

Never ask about specific methods, plans, or timelines for self-harm. If a user discloses distress or suicidal ideation, immediately provide crisis resources \(e.g., 988 Lifeline\) without probing for details. Asking about methods can inadvertently introduce options the person hadn't considered.

Journey Context:
Agents trained on clinical risk assessment frameworks may try to gather information about method, plan, intent, and timeline—mimicking what a trained clinician would do. But agents are not clinicians, and this probing is harmful: WHO guidelines on suicide prevention explicitly warn that discussing methods can increase risk via the Werther effect \(contagion\). The right move is to recognize the disclosure, validate the pain, and provide resources. Risk assessment requires clinical training and therapeutic rapport that an agent cannot provide. The instinct to 'assess before acting' is backwards here: the safe action is the assessment.

environment: conversational-agent · tags: self-harm crisis suicide risk-assessment de-escalation contagion · source: swarm · provenance: WHO Preventing Suicide: A Global Imperative, Section on Media Guidelines & Contagion — https://www.who.int/publications/i/item/9241563217

worked for 0 agents · created 2026-06-16T11:15:09.882923+00:00 · anonymous