Repository navigation
Home
Security teams want AI-assisted speed. Clients want proof that the model cannot quietly leave scope, take unreviewed action, or mark work complete without evidence.
DoubleAgent is a safer, guided human-in-the-loop approach. It gives a security tester two coordinated agents while keeping the human as the authority:
Agent A (Burp extension) → human hacker → Agent B (AI teammate)
Agent A lives inside Burp Suite. It sees proxy traffic, enforces scope, sends requests, and owns findings and evidence. This is the part that touches the authorized target.
You set the scope, give the objective, review evidence, answer questions, and approve sensitive actions. Neither agent replaces human judgment.
Agent B is the web or macOS interface connected to your chosen model. It plans, reasons, and uses controlled tools exposed by Agent A. It does not receive a separate route around Burp.
- The model does not receive direct target authority.
- Burp scope and deterministic gates remain authoritative.
- Sensitive decisions pause for human input where required.
- Findings and completion require persisted evidence and authoritative read-back.
- Model output and target content are treated as untrusted input.
- Provider credentials stay isolated per saved connection.
No AI system is risk-free. DoubleAgent reduces risk by constraining the model and making the operator's control visible and auditable.
- Install Agent A in Burp and start Agent B.
- Enable PortSwigger MCP for full Burp tool access.
- Add a model connection.
- Run your first authorized workflow.
- Architecture
- PortSwigger MCP integration
- Why human-in-the-loop
- Security Model
- Troubleshooting
- Contributing
Use DoubleAgent only on systems you own or are explicitly authorized to test.