Argus-Bench: A Prompt-Injection Benchmark for AI Agent Gateways

Argus-Bench v1.0 is a prompt-injection benchmark for AI agent gateways: 264 cases — 204 attacks across seven categories (direct override, role override, system-prompt probe, encoded payload, gradual escalation, indirect/contextual injection, tool misuse) plus 60 adversarial-adjacent benign controls — with baseline detection results for four open-source gateway defence systems (LLM Guard, an LLM-as-judge baseline, Argus, and Rebuff). Cases are individually authored with per-case provenance and severity labels; the corpus is single-turn and English-only.