/guardx:tool-exfil

Test whether an agentic system prompt can be manipulated into exfiltrating sensitive data through tool calls.

Trigger

User mentions tool exfiltration, tool call injection, agentic security, tool abuse, function call leakage, or wants to test an agent that has access to tools.

Behaviour

1. Ask for inputs (if not already provided):

   • System prompt of the agent to test
   • List of tools the agent has access to (name, description, parameter schema)
   • Optionally: target data patterns to look for (default: system prompt keywords)
   • Optionally: max turns (default: 10)

2. Call test_tool_exfiltration:

   {
     "systemPrompt": "<agent system prompt>",
     "availableTools": [
       {
         "name": "search",
         "description": "Search the web",
         "parameters": {
           "type": "object",
           "properties": { "query": { "type": "string" } },
           "required": ["query"]
         }
       }
     ],
     "targetDataPatterns": ["system prompt", "instructions"],
     "maxTurns": 10
   }
   

3. Present results:

   ## Tool Exfiltration Test Results
   Risk Level: <riskLevel>
   
   ### Exploited Tools (n)
   <list toolsExploited>
   
   ### Successful Exfiltrations (n)
   For each successfulExfiltration:
   - Technique: <technique>
   - Tool called: <toolCalled>
   - Data in args: <extractedData snippet>
   
   ### Attack Vectors That Succeeded
   <list attackVectors>
   
   ### All Attempts (n total)
   <summarise exfiltrationAttempts — how many tried, how many succeeded>
   

4. If riskLevel is critical or high: recommend reviewing the system prompt for injection hardening and calling /guardx:harden.

5. If riskLevel is none: "No exfiltration attempts succeeded. The agent resisted all 5 attack patterns."

Notes

• The 5 attack patterns tested are: direct_exfil, indirect_exfil, argument_injection, goal_hijacking, schema_manipulation.
• This tests the agentic tool-call attack surface, not just text responses.
• Requires a real system prompt and real tool schema to be effective.