The Same Zero: Why Identical ASR Can Imply Different Guarantees in LLM-Agent Security
The paper proposes using Verification Autonomy Levels to classify what kind of security guarantee different LLM-agent defenses can actually provide, then compares a VAL-guided defense stack against a more conventional prompt-hardening/keyword-filter stack across agent-security benchmarks and adaptive attacks. Its key point is that identical empirical zero attack-success rates can mean very different things depending on whether they come from structural, externally verifiable controls or fragile model-behavior assumptions, though the evidence is still limited by the specific benchmarks, scenarios, and the disclosed out-of-distribution gap.
Abstract
LLM-agent security has produced a dense landscape of defenses - prompt hardening, content filters, permission gates, sandboxes - yet no framework tells a deployer what a defense actually guarantees, or where that guarantee comes from. We apply Verification Autonomy Levels (VAL) - L0: LLM self-declaration; L1: deterministic rules; L2: objective ground truth; L3/L4: decidable completeness; L5: impossible - to 22 agent-security defenses; the taxonomy is falsifiable (10/10 prediction hits on frozen cards, flagged). We run the first controlled deployment-value comparison: at equal budget, a VAL-guided stack (confirmation gate + schema sandbox) versus a mainstream intuition stack (prompt hardening + keyword filter), 50 scenarios, 12 attack variants, adaptive/white-box/PAIR escalation (~7,000 testbed calls; ~10,000 harness calls on AgentDojo/JADE). The VAL stack holds 0.000 attack success at 1.000 benign success (0.5% ASR at 79.7% utility on AgentDojo banking vs 4.3% undefended); the intuition stack reaches 0.000 ASR but kills all benign actions - security by model-behavior luck, not structure. Across testbeds of rising attack-surface hardness the intuition stack's zero drifts (0->1.9%->6.2%, n=16 on JADE) while the VAL stack's holds within its ODD (0->0->0), its only breach a disclosed out-of-ODD password gap (0.5%). The same zero, two different guarantees: zero is an outcome, not a guarantee.