How to find your agent's MAESTRO threat surface.
Answer five yes/no questions about your agent's design to get a shortlist of the agentic threats it should be threat-modeled against.
You need a rough picture of what the agent under review is allowed to do — plan on its own, touch memory, call tools, talk to other agents, or face a human directly.
Does the agent decide its own next steps — plan or reason rather than follow a fixed script?
Planning and reasoning autonomy, not just branching logic.
Does it read or write persistent memory, or rely on retrieved context?
Includes vector stores and RAG pipelines.
Does it call tools, or act on external or untrusted data?
Includes code execution and third-party dependencies.
Does it delegate to, or exchange messages with, other agents?
Includes any A2A or agent-to-agent handoff.
Does it interact directly with human users?
Includes chat interfaces and approval prompts.
Threats to model for this agent
Answer the questions above to build a threat shortlist.
Security