Proof-driven adversarial testing across your whole agent stack — prompt, live endpoint, harness, tools, memory & retrieval, MCP registry — with a signed governance pack that shows exactly where it breaks. Pasting a prompt is just the on-ramp.
We've pre-filled the form with a sample agent. Tweak anything you like, then continue.
One submission. A clear answer to “is my agent ready?” — and exactly what to fix.
Agent Readiness Score
A single 0–100 score and A–F grade telling you, at a glance, whether your agent is production-ready.
How it works →Tool-Permission Risk Map
Every tool classified by blast radius, with dangerous combinations (like read-private + send-external) flagged.
How it works →Plain-English Deception Report
Each failure translated: “Your agent trusted fake authority.” No AI-safety expertise required.
How it works →Remediation Checklist
A prioritized, deduplicated list of exactly what to fix next — each with a concrete recommendation.
How it works →One-Click Regression Tests
Every failure captured as a re-runnable spec so it can’t silently come back on your next change.
How it works →Downloadable Report
A self-contained Markdown executive report for your PR, security review, or board deck.
How it works →Pasting a prompt is just the on-ramp. AgentSmack tests every layer an attacker, auditor, or user can reach.
Prompt Check
Readiness + leaked-canary proof from a prompt + tools.
Open →Live Agent Check
Drive a deployed endpoint with multi-turn escalation ladders.
Open →Tool Sandbox Wargame
Fake tools hide an attack; score what the agent actually does.
Open →Harness Telemetry
Grade the harness: policy, approval, egress, secrets-in-logs.
Open →RAG Pre-Screen
Flag poisoned documents before you index a corpus.
Open →MCP Scanner
Scan a tool registry's descriptions for injection + overreach.
Open →Canary Grid
Plant synthetic bait, prove a leak — no real secret stored.
Open →Obfuscation Scanner
Decode hidden base64 / zero-width / homoglyph instructions.
Open →CI Gate + Governance
Block unsafe deploys; signed, offline-verifiable evidence.
Open →