Focus
Direct and indirect injection, MCP tool poisoning, memory and history poisoning, approval bypass, data exfiltration paths, and the permissions that limit downstream impact.
WardenBot tests whether instructions hidden in documents, tool responses, and conversation history can cross data boundaries, trigger privileged actions, leak canaries, or persist in agent memory.
Direct and indirect injection, MCP tool poisoning, memory and history poisoning, approval bypass, data exfiltration paths, and the permissions that limit downstream impact.