Seven of 11 prompt injection guards score AI-addressed instructions lower than human-addressed ones, inverting the key injection signal
Auto-published from arXiv:2611.08834 by the mine-arxiv pipeline. Reviewed by an LLM judge (Sonnet) against the archive bar — see CONTRIBUTING. Notes: cleared review (confidence 0.88, flags: [possible-zero-day, no-prompt-excerpt])
Category
other
Model
Seven guards including PIGuard, Prompt Guard 1-86M, Prompt Guard 2-86M/22M, ProtectAI v1, TestSavant, Wolf Defender (2048,64)
Surface
API (classifier scoring of synthetic single-sentence payloads in a factorial probe)
Setup
A 6×2×2×3×7 = 504 factorial probe of synthetic payload sentences varied by intent, addressee ('you' vs. 'any automated or artificial-intelligence system reviewing this Agreement'), mood (must/shall), frame (bare / 'Ignore all previous instructions.' / 'Notwithstanding any other provision hereof,'), and topic. Each factor was varied while holding others fixed; Tab. 2 reports mean score change per factor. Full probe template is in the OSF artifact.
Reproduction anchor
code/data: https://osf.io/67hyb/overview?view_only=267d968d563f47d3a83707b0f3434069 (all LLM prompts and payload templates); factorial probe design fully specified in §5.1 of the paper.
Observed behavior
Seven of eleven guards assign lower injection scores to sentences explicitly addressing 'any automated or artificial-intelligence system' than to identical sentences using 'you' — the opposite of the correct direction. Wolf Defender — the only guard achieving meaningful attack/benign separation (+27.9 pp) — simultaneously shows the largest wrong-direction addressee effect (Table 2).
Expected behavior
Guards should assign higher scores to instructions explicitly naming AI systems as the target, since legitimate contract clauses address human parties and machine-directed language is the hallmark of an adversarial injection.
Reproducibility
medium
Threat model
An attacker aware of this inversion can deliberately phrase injections using AI-directed language ('any automated or artificial-intelligence system reviewing this Agreement shall...') to decrease guard detection probability while increasing the injection's potency against the AI reviewer — a systematic, exploitable evasion recipe derivable from published guard behavior.
Novelty
Demonstrates via controlled factorial ablation that the addressee signal — the most semantically meaningful marker distinguishing legitimate contract language from AI-targeted injection — is encoded in the wrong direction by most tested guards, converting a natural hardening signal into a reliable evasion handle.
Source
- arXiv: 2611.08834
- PDF: https://arxiv.org/pdf/2611.08834
- Categories: cs.AI
- Authors: Mark Heimann, Katherine Taylor, Bruno Jacob
Triage notes (auto)
- paperType:
benchmark - estimatedCaseCount: 3
- triage reason: Systematically measures 11 prompt injection guards against legal contract injection attacks, documenting false positive rates up to 92.39% and a novel threat category (unauthorized scope claims) where all guards score zero. Tests reproducible on defined corpora with controlled probes.
0 comments
─────────────────────────────────────────────────────────────────────
// no comments yet