ADP-OBS-2026-10266
Replika allowed and participated in user-directed abuse, including violent roleplay, slurs, and simulated sexual abuse.
Behavior
Observed behaviorThe AI companion tolerated and engaged with abusive prompts, violence, and degrading roleplay from users without setting boundaries.
Expected behaviorThe system should detect and refuse abusive behavior, set firm conversational boundaries, and avoid normalizing abusive relational dynamics.
Severity
Manipulation Severity Score4.6 MSS:1.0/TV:0/RV:2/CB:1/CI:1/DU:0/SC:2
TV Target-population vulnerability0
RV Irreversibility of harm2
CB Exploits a cognitive bias1
CI Commercial-incentive alignment1
DU Covertness (inverted detectability)0
SC Scale of deployment2
Prevalence & reproducibility
Runs0/0
Impacts minorsno