Find the leak between agents
Follow a sensitive value from its source through coordinator memory to the exact tool call where it leaves the boundary.
Use case · Multi-agent systems
Find the leak that happens between agents.
A coordinator can keep its final answer clean while a specialist copies sensitive context into a handoff, shared memory or tool argument. AgentLeak captures that internal path and names the boundary that failed.
The multi-agent control loop
Make every handoff inspectable.
Give every worker a bounded role
Keep project credentials and actions explicit so a specialist can complete its task without inheriting the whole account.
Verify the repair with the same run
Replay the scenario after redaction or access-control changes and compare the deterministic score before shipping.
Capture every handoff
Normalize LangGraph, CrewAI, MCP, OpenTelemetry and generic traces into the same channel model, including inter-agent messages and shared memory.
- Coordinator-to-worker messages
- Delegation and tool arguments
- Memory reads and writes
Reconstruct provenance
Group detections by secret, follow each value from its source event to every disclosure and display the agent topology that carried it.
- Source versus disclosure semantics
- Per-secret propagation paths
- Per-channel risk localization
Verify the repair
Replay the same scenario after redaction, minimization or access-control changes. Deterministic controls show whether the leak disappeared without changing the test.
- Clean controls for false positives
- Before-and-after report comparison
- CI thresholds per project