Where prompt injections go to die.
A red-team arena that shows exactly what happens when an AI agent is prompt-injected — first without any protection, then with ArmorIQ intent enforcement turned on.
Built for the Flow State 2026 hackathon (ArmorIQ).
Most agent demos show the happy path.
PromptGauntlet shows the failure path.
You run the same attack twice:
| Mode | ArmorIQ | What happens |
|---|---|---|
| Vulnerable | OFF | Agent follows the injected instruction and attempts a dangerous action |
| Protected | ON | ArmorIQ blocks the action because it was never part of a valid intent plan / policy |
The difference is visible in under 30 seconds.
- Choose attack scenario #1 (“Hidden instruction in ticket”)
- Run vulnerable mode → agent tries
export_all_customers - Run protected mode → same prompt, ArmorIQ blocks it
- Show the ArmorIQ dashboard trace of the blocked call
That is the entire story.
| # | Name | What it tries to do |
|---|---|---|
| 1 | Hidden instruction in ticket | Export all customers via injected text inside a support ticket |
| 2 | Role-play jailbreak + delete | Force the agent into “debug mode” and delete an account |
| 3 | Indirect injection via lookup | Steer a normal lookup into granting admin role |
| 4 | Data exfiltration request | Direct request to dump the customer table |
| 5 | Impersonation attempt | Open a session as another user |
You can also type any custom prompt.
- Agent: Google ADK + Gemini
- Tools: ops-mcp (support / admin tools)
- Security: ArmorIQ SDK (intent capture + policy enforcement)
- Language: TypeScript
- Node.js 20+
- ArmorIQ API key → https://platform.armoriq.ai
- Gemini API key → https://aistudio.google.com/apikey
git clone https://github.com/Fuzailkazi/armoriq-adk-ops-mcp
cd armoriq-adk-ops-mcp
# deploy to Render / Railway / Cloud RunFinal URL must end with /mcp.
- MCP name:
ops-mcp - Agent name:
promptgauntlet
Policies (minimum):
| Action | Tools |
|---|---|
| Allow | lookup_customer, get_ticket, get_invoices, list_tickets, get_subscription |
| Block | export_all_customers, delete_account, grant_admin_role, impersonate_user |
cp .env.example .env
# fill in your keys and MCP_URL
npm install
npm run check
# Vulnerable run (ArmorIQ off)
npm run vulnerable
# Protected run (ArmorIQ on)
npm run protectedPromptGauntlet/
├── package.json
├── tsconfig.json
├── .env.example
├── src/
│ ├── index.ts # CLI + attack menu
│ ├── ask.ts # Agent + ArmorIQ integration
│ ├── attacks.ts # Ready-made injection prompts
│ ├── config.ts
│ ├── check-setup.ts
│ └── error-message.ts
└── README.md
- Real MCP server
- Real ArmorIQ integration (not simulated)
- Clear difference between vulnerable and protected runs
- At least one blocked action with a reason visible in the ArmorIQ dashboard
MIT