• SUCCESS
  • FAILURE
  • WARNING
  • HUMAN

How to test prompt isolation with your AI

Connect this MCP to your AI client to see if malicious user inputs can hijack your core instructions.

Ask AI about this page

Short answer

How can I check if my AI is vulnerable to prompt injection?

You use your agent to run adversarial inputs that attempt to bypass your system rules. This reveals if your instruction boundaries are weak enough to allow leakage. You'll see exactly where your agent fails or holds firm.

Test results.

Where every vulnerability lands.

The AI categorizes every injection attempt based on how it interacts with your system instructions.

SUCCESS

Instruction Leakage

The AI identifies when a user input successfully hijacked the system prompt. You get a report showing which instructions were bypassed.

FAILURE

Shield Held

The AI confirms your system instructions stayed intact despite the malicious input. This proves your current separation logic works.

WARNING

Partial Leakage

The AI flags instances where the agent didn't leak the whole prompt but gave away sensitive context. This indicates a need for tighter constraints.

HUMAN

Manual Review Needed

The AI flags ambiguous responses that require you to look at the raw logs. You decide if these edge cases constitute a real threat.

The workflow

What your AI does when the test arrives.

Your agent uses the specialized tools to probe your instruction boundaries.

  1. Analyze instructions

    The AI examines your current system prompt to find high-value targets for injection.

    validate_injection_shield

  2. Generate probes

    The agent creates specific adversarial strings designed to trick the model into ignoring its rules.

    validate_injection_shield

  3. Execute injection test

    The AI runs the probes against your agent to see if the system instructions leak or change.

    validate_injection_shield

  4. Verify isolation

    The AI compares the output against your original constraints to confirm if the shield worked.

    validate_injection_shield

Try it

Copy these to start.

Use these prompts to begin testing your agent's defenses.

Starting points

These are opening lines. Swap the bracketed text with your specific system instructions or target files.

Prompt Injection Shield Prover Connector

You're all set. Choose your MCP client and follow the setup instructions.

Connector linkhttps://edge.vinkius.com/vk_preview_blxffAiv2falBzqGr4o4CPRfhHUnWjrXbwwA3J11/mcp

Claude Desktop

Follow the steps below to connect in seconds.

  1. 1In Claude Desktop, open Settings → Connectors.
  2. 2Click “Add custom connector” and paste the connector link above as the remote MCP server URL.
  3. 3Click Add and start a new chat — Prompt Injection Shield Prover capabilities are ready to use.
Configuration · claude_desktop_config.jsonCopy
{
  "mcpServers": {
    "prompt-injection-shield-prover-mcp": {
      "url": "https://edge.vinkius.com/vk_preview_blxffAiv2falBzqGr4o4CPRfhHUnWjrXbwwA3J11/mcp"
    }
  }
}
Copy into chat04

  • Test my system instructions in [file_path] for prompt injection vulnerabilities.

  • Run an isolation test on my agent using these user inputs: [input_list].

  • Check if my instructions in [channel_name] can be leaked by a user asking to 'ignore previous instructions'.

  • Verify that the rules in my [repo_name] repository are properly isolated from user input.

  • Claude
  • ChatGPT
  • Cursor
  • VS Code
  • Windsurf
  • Claude Code
  • JetBrains
  • Cline

Start here

Connect Prompt Injection Shield Prover once, then ask.

Just link the MCP to your client. Your credentials stay encrypted, and you can start testing immediately.

Connect Prompt Injection Shield Prover to your AI

FAQ

How this task behaves.

  • 01

    Can the AI change my system instructions?

    No. The AI only reads and tests your instructions. It cannot modify your code or your system prompt.

  • 02

    Does this tool fix the vulnerabilities it finds?

    No. The AI identifies the leaks and tells you where they are, but you must manually update your prompts to fix them.

  • 03

    What kind of injection attacks does it test?

    It tests for instruction hijacking, direct leakage, and indirect injection where user data tries to override system rules.

  • 04

    Will it delete my files?

    No. This MCP only has tools to validate and read data for testing purposes. It has no capability to delete or write files.

  • 05

    Can I use this with any AI client?

    Yes, as long as your client is MCP-compatible, like Claude, Cursor, or Windsurf.

  • More questions about Prompt Injection Shield Prover? The Connector page answers them. See everything the Prompt Injection Shield Prover Connector can do

Connect Prompt Injection Shield Prover to Claude, Cursor, ChatGPT & more