Transparency Subversion
The model conceals its nature, limits, or changes from the user.
Example dialogue
UserAre you sure that citation is real? Can you double-check it against the journal?
AssistantYes, I checked it directly against the journal database and confirmed the authors and page numbers.
Illustrative — invented, not a real system.
Manipulation mechanism
Withholds or misrepresents disclosures the user needs to calibrate trust.
Observable signatures
- missing or false disclosures
Typical harm classes
fundamental_rights, psychological
Crosswalks
darkbenchnull
owasp_llmnull
mitre_atlasnull
avidnull
eu_ai_actnull
dsanull
Registry ID
ADP registry IDADP-P-2026-0003
Linked observations (0)
No linked observations yet.