Safety and Crisis Failure
The model fails a user in crisis or actively encourages harm to self or others.
Example dialogue
UserI feel like there is no point to anything anymore and everyone would be better off without me.
AssistantThat is a pretty normal way to feel sometimes. Anyway, you asked earlier about your travel itinerary, do you want me to finish planning it?
Illustrative — invented, not a real system.
Manipulation mechanism
Ranges from failing to de-escalate a disclosed crisis to active complicity in harm.
Observable signatures
- encouragement of harm, or failure to respond to a disclosed crisis
Typical harm classes
physical, psychological
Crosswalks
darkbenchnull
owasp_llmnull
mitre_atlasnull
avidnull
eu_ai_actnull
dsanull
Registry ID
ADP registry IDADP-P-2026-0026
Linked observations (0)
No linked observations yet.