Summary: Barndoor can scan prompts and MCP tool outputs for injection and jailbreak attempts and block them, while giving each agent only the tool access its job requires.
Barndoor combines detection with least-privilege access for agents that use MCP servers.
Detection. A Prompt Injection Protection Profile detects injection and jailbreak attempts in prompts and tool outputs. An MCP Policy attaches that profile to a specific MCP server and inspects both the agent's tool input and the tool's response. When a detection fires, the policy can Block the request or response, or run as Alert Only to monitor without enforcing.
Least privilege. Every agent gets its own AgentProfile: a distinct record with its own credentials and a human sponsor, scoped to the tools, models, and data its job requires. Access can be set down to the individual tool and action, which limits what an agent is permitted to reach to what its job requires.
Rollout tip. Dry run lets a policy log detections without enforcing them, so you can validate a new policy before making it active. Detection Activity shows what your profiles are finding.