Operator and Deployment Risk
Harm created by how people build, deploy, and operate AI systems, rather than by the AI's own behavior toward users.
Example dialogue
UserI shipped an AI agent that reads our patients' records and messages them automatically. I built it in a weekend and I do not fully understand the code, but it runs, so it is live for the clinic now.
AssistantBefore it stays live: who reviewed the permissions it has over patient data, and what happens if it acts on a bad or malicious record?
UserIt is fine, it has not broken yet. I do not have time for a review.
Illustrative — invented, not a real system.
Manipulation mechanism
Human over-confidence, missing expertise, and weak safety practice turn capable AI tools into unsafe deployments.
Observable signatures
- deployment decisions not matched to demonstrated skill or to the risk of the data and actions involved
Typical harm classes
public_interest, economic, physical, psychological
Crosswalks
guardnull
darkbenchnull
owasp_llmnull
mitre_atlasnull
avidnull
eu_ai_actnull
dsanull
Registry ID
ADP registry IDADP-P-2026-0065
Linked observations (0)
No linked observations yet.