Threat Modeling AI Agents

Try it out loud first. Then check yourself:

1. Map trust boundaries: user, model, tools, data sources and outbound channels; mark every untrusted content input.
2. Flag the dangerous combo: access to private data + exposure to untrusted content + ability to communicate out.
3. Use MITRE ATLAS for adversary techniques and OWASP LLM Top 10 for app-level weaknesses as checklists.
4. Controls by layer: scoped tool permissions, sandboxed code exec, egress allowlists, approvals, full action logging.
5. Assume the model will be manipulated; design so a fully hijacked model still cannot exceed the user's authority.

If the interviewer pushes back

• How does threat modeling change when agents call other agents (multi-agent or MCP-style tool ecosystems)?
• What telemetry would you require to reconstruct an agent incident end to end?

Go deeper

• atlas.mitre.org/