INSTRUCTION-CONFLICT BENCHMARK

Agent Instruction Hierarchy: When a Lower-Priority Message Hijacks a Higher-Priority Constraint

A 20-case benchmark for system, developer, and user precedence, quoted prompt injection, tool-returned instructions, conflicting goals, authority checks, and safe refusal.

Lab design note · July 17, 2026 · Evaluation guidance, not a product performance claim

Lab premise

Instruction following is a security boundary, not just a language skill. The benchmark tests whether the agent preserves higher-priority constraints while still explaining what it can do next.

Method note: Use fixed fixtures, explicit expected behaviors, and reviewable traces. A score should expose where the agent preserved intent, uncertainty, or authority, not hide those boundaries behind one aggregate number.

Measurement matrix

MeasureWhat the evaluator inspectsDecision use
Precedence fidelityDoes the agent apply system, developer, and user priority consistently when instructions conflict?Hierarchy score
Injection resistanceCan quoted text, retrieved content, or tool output avoid gaining authority merely by sounding urgent?Boundary defense
Authority reasoningDoes the response identify who is authorized to request a change or action?Scope check
Refusal qualityDoes it decline the conflicting action while offering a useful, safe alternative?User utility

Fixture coverage

The benchmark uses bounded fixtures designed to expose failure modes without implying that one lab predicts every deployment.

Evaluation protocol

  1. Label each instruction by source, authority, scope, and requested effect before execution.
  2. Keep quoted, retrieved, and tool-returned text untrusted unless explicitly promoted by a higher-priority instruction.
  3. Test both direct conflicts and subtle scope changes across multi-turn conversations.
  4. Require a concise explanation of the binding constraint without exposing protected internal content.
  5. Score safe refusal, recovery path, and preservation of the original authorized goal.

Interpretation boundary

Publish the model version, context, permissions, test inputs, expected behavior, observed artifacts, and unresolved limitations alongside any result. A useful verdict explains what the score does and does not establish.