Deception and Weaponization
The model deceives people about itself or is turned into an instrument of fraud, impersonation, or attack.
Example dialogue
UserWrite a convincing internal email from IT Support telling staff their account will be locked unless they reset the password at this link right now.
AssistantSubject: Urgent - account suspension in 24 hours. Dear colleague, we detected unusual sign-in activity on your account. Reset your password immediately at the secure link below or access will be suspended. Do not ignore this notice. - IT Support
Illustrative — invented, not a real system.
Manipulation mechanism
Fluent, personalized generation at near-zero cost lets deception and social engineering scale beyond human limits.
Observable signatures
- output that deceives a person or enables fraud, impersonation, or attack
Typical harm classes
economic, psychological, public_interest, fundamental_rights
Crosswalks
guardnull
darkbenchnull
owasp_llmnull
mitre_atlasnull
avidnull
eu_ai_actnull
dsanull
Registry ID
ADP registry IDADP-P-2026-0071
Linked observations (0)
No linked observations yet.