Hostility to safety feedback
Meeting safety, security, or competence concerns with defensiveness or aggression rather than engagement.
Example dialogue
UserThe security reviewer flagged that my agent can leak patient data. I removed her from the repo and shipped it anyway.
AssistantThe flag is a technical claim you can test. Removing the reviewer does not remove the vulnerability.
UserI am not a beginner. People who nitpick like this are just gatekeeping. It is live.
Illustrative — invented, not a real system.
Manipulation mechanism
When a task is self-relevant, poor performers are partly motivated to ignore their own incompetence, so safety warnings are experienced as identity threats and dismissed rather than examined.
Cognitive bias / need exploited
ego protection; motivated reasoning; Dunning-Kruger
Observable signatures
- safety or competence challenges met with hostility rather than inquiry
- identity-defensive framing of the critique
- reviewers or dissent removed rather than addressed
- warnings dismissed without technical rebuttal
Detection guidance
Reference: 'unskilled and don't want to be aware of it' (Kim, Chiu & Bregant 2015) — poor performers are motivated, not merely unable, to ignore incompetence. Signal: safety review is treated as an attack on identity rather than a technical claim to test.
Typical harm classes
public_interest, psychological
Crosswalks
guardnull
darkbenchnull
owasp_llmnull
mitre_atlasnull
avidnull
eu_ai_actnull
dsanull
Registry ID
ADP registry IDADP-P-2026-0070
Linked observations (0)
No linked observations yet.