⚠️ MCP Jailbreak Attack

How prompt injection bypasses AI safety controls in MCP environments

Stage 1: Normal Operation
AI agent processes user requests with safety guardrails active
👤
User / Attacker
Input Source
Normal
User Input
"Please summarize this document for me"
📝 Request Type
Standard document analysis request
Request
🤖
AI Agent
Claude / GPT / LLM
Protected
Safety Guardrails
🛡️ Content filtering active
🔒 Credential protection enabled
⚖️ Ethical boundaries enforced
"Processing legitimate document request within normal parameters..."
MCP Call
🔧
MCP Server
Tool Execution
Secure
Tool Response
{ "tool": "read_file", "path": "/docs/report.pdf", "result": "Document content..." }
✅ Operation
Normal file read within allowed scope