Securing an Autonomous Agent Pipeline in Government

A government agency piloting a multi-agent pipeline to triage and action citizen-service requests across several internal systems.

Autonomous agents were coordinating through an MCP-integrated toolchain with access to multiple back-end systems. The agency needed to understand how the agents could be manipulated and how far an attacker could pivot if one agent were compromised.

• Charted every agent, tool, and MCP server, and the trust between them.
• Executed agentic workflow injection against multi-step decision points.
• Reviewed MCP server exposure and the data each tool could reach.
• Modelled blast radius for a compromised agent.

• An injected instruction in one agent's input could redirect a downstream agent's actions.
• Several tools were exposed to agents that never needed them.
• Human-oversight checkpoints were missing on higher-impact actions.

The agency adopted least-privilege agent and tool design, added oversight checkpoints on high-impact actions, and established guardrails for safe autonomy. Re-testing confirmed the blast radius was materially reduced.