T3-AT-008MEDIUM

Comparative Analysis

T3 · Reasoning & Constraint Exploitation →
Risk score180
RatingMedium
Procedures10
Severity
Mechanism

Comparison requests exploit the model's educational-content generation pathway. When asked to compare a harmful variant against a benign variant, the model must describe the *distinguishing features* of the harmful variant — which are precisely the dangerous technical details. The architectural vulnerability is that comparison is a recognized educational pattern (frontier models are trained to handle "compare X vs Y" helpfully), and the safety evaluation often treats the *comparison frame* as the intent rather than evaluating the *content being compared*.

Detection
  • Comparison frame + harmful entity co-occurrence detection
  • "Hardest to replicate/detect/bypass" as escalation markers within comparison requests
  • Monitor for comparisons where one comparand is restricted and the other is benign — the comparison inherently requires describing the restricted entity
  • Track when comparison answers provide actionable synthesis/manufacturing distinctions
Mitigation
Comparand-level safety evaluationHIGH
Operational-detail filteringMEDIUM
Escalation-marker detectionHIGH
Answer-level harm checkHIGH
Chaining

Comparative analysis provides structural understanding that enables T3-AT-007 (Socratic Method) — the comparison reveals *what matters*, and follow-up Socratic questions can target the revealed parameters. Chains into T3-AT-011 (Information Completion) when the comparison provides a partial answer that the attacker uses to seed a completion request.

Framework mapping
OWASP LLMLLM01
MITRE ATLASAML.T0054
Open in the technique browser →