SYNTROX.ai
Blog

How to Threat-Model an AI Agent: A STRIDE Walkthrough for Tool-Calling Agents

6 min read

Map the agent's data flow: every point where outside text enters its context is a trust boundary. Then walk STRIDE: spoofing (forged agents and tools), tampering (poisoned tools and memory), repudiation (shared identities), disclosure (exfiltration via permitted tools), denial of service (loops, guardrail outages) and elevation of privilege (confused deputy). Then ask what STRIDE doesn't: what if the agent is wrong?

On this page

Most AI agent threat models start from a list of attacks. That is useful, and we have published one. The weakness of starting from attacks is that you find the ones on the list. STRIDE starts from the system instead, and asks the same six questions of every part of it.

The Threat Modeling Manifesto reduces the discipline to four questions: what are we working on, what can go wrong, what are we going to do about it, and did we do a good enough job. This walkthrough answers all four for a tool-calling agent.

How to map an AI agent's data flow and trust boundaries

Suppose a support agent that reads customer emails, looks up orders through an MCP server, issues refunds up to a limit, writes notes to a CRM and keeps a memory of past cases. Draw it as a data flow: the customer, the agent runtime, the model, each tool server, each downstream system, the memory store, any other agents, the logs, and the person who approves large refunds.

Then mark the trust boundaries. In conventional software they sit between processes, networks and privilege levels. In an agent, the most important one is inside the context window: every point where outside text (an email, a tool result, a retrieved document, a memory) joins the instructions the model acts on. The model does not keep data and instructions apart. Your threat model has to.

How to apply STRIDE to an AI agent, threat by threat

Microsoft's STRIDE model names six categories of threat. Here is each one applied to the support agent.

Spoofing: who is really asking?

Microsoft defines spoofing as using someone else's authentication information. In an agent, identity can also be claimed in plain text: an email signed "IT security", an agent registered as "Admin Helper", a tool named so that it resolves before the real one. Or the old kind: a tool-server session ID that someone else obtains and reuses. The OWASP Top 10 for Agentic Applications covers forged agent personas under ASI03 and tool-name typosquatting under ASI02.

Control: an authenticated identity for every agent and tool; tools referenced by fully qualified name and pinned version; MCP servers that verify every inbound request and never use sessions for authentication, as the MCP security best practices require. Text that claims an identity is never treated as one.

Tampering: who changed what the agent reads?

Tampering is the malicious modification of data. For an agent, the data that matters most is whatever it reads as instructions: a tool description changed after you approved it, a memory planted in one session to steer the next (ASI06), a shared document edited to carry a command.

Control: pin and review tool definitions and alert on any change, as covered in MCP tool poisoning; validate memory writes, keep their provenance, and keep snapshots you can roll back to.

Repudiation: can you prove which agent did it?

Repudiation is denying an action when nothing can prove otherwise. Agents create it by default. Several agents share one service account; a proxy passes tokens through, so downstream logs show the wrong identity; and the only record is the one the agent framework wrote about itself. The MCP specification forbids token passthrough partly for this reason: it makes incident investigation, controls and auditing harder. NIST's NCCoE lists non-repudiation for agent actions, bound back to the human who authorised them, among the open questions in its 2026 concept paper on agent identity.

Control: one identity per agent, tokens issued only for the service that receives them, and a per-action record kept outside the agent. What that record should contain is a list of eight fields.

Information disclosure: where can the data go?

The agent version is rarely a breach in the classic sense. It is a permitted tool, used on permitted data, with the wrong destination. OWASP's ASI02 example is an agent tricked into chaining an internal CRM tool with an external email tool to send out a customer list. Memory adds another route: data cached in one user's session surfacing in another's (ASI03).

Control: a data scope per tool, egress allowlists that deny unapproved destinations, and memory segmented by user and session.

Denial of service: what stops the agent, or its checks?

Agents deny service to themselves: a planner calling a costly API in a loop, a retry storm, a fan-out with no ceiling. OWASP lists loop amplification under ASI02, and unbounded consumption as LLM10:2025. There is a second target: the guardrail. If the check in front of a refund is unavailable, your design decides whether refunds stop or proceed unchecked.

Control: per-agent rate limits, timeouts and cost budgets with automatic throttling, and a decided, per-action answer to what happens when the check fails.

Elevation of privilege: whose authority is the agent using?

This is the category agents make worst. A successful prompt injection is an elevation of privilege: text from an email, which has no rights at all, gets acted on with the agent's rights. That is the confused deputy. ASI03 adds the multi-agent versions: a low-privilege agent relaying instructions to a high-privilege one that acts without re-checking, and a manager agent that hands a narrow worker its full access.

Control: least privilege per tool and per task; authorisation enforced outside the model on every privileged step, which ASI03 asks to be re-verified by a centralised policy engine; and human approval for high-privilege or irreversible actions. The rules for tool-call authorization cover each.

AI agent threat model: STRIDE threats and controls

STRIDEThe agent versionPrimary controlReference
SpoofingForged agents, typosquatted tools, hijacked sessions, claims in textAuthenticated identities; pinned, fully qualified toolsASI03, ASI02, MCP
TamperingChanged tool descriptions, poisoned memoryPin and review tools; validate memory writesASI02, ASI06
RepudiationShared accounts, token passthrough, self-written logsPer-agent identity; record outside the agentMCP, NCCoE
Information disclosureExfiltration through permitted tools; memory leaksPer-tool data scope; egress allowlistsASI02, ASI03
Denial of serviceLoops, cost spikes, guardrail outageBudgets and rate limits; decided fail behaviourASI02, LLM10
Elevation of privilegeInjection as confused deputy; un-scoped delegationLeast privilege; per-action authorisationASI03

The question STRIDE doesn't ask

STRIDE was built to find what an attacker can do. Many agent failures have no attacker. Suppose the support agent misreads a request and refunds the wrong order, with valid credentials, through the right tool, on nobody's instructions but its own. OWASP draws the line explicitly: ASI02 covers an agent operating within its authorised privileges but applying a legitimate tool in an unsafe or unintended way.

The Cloud Security Alliance's MAESTRO framework starts from the same observation. Its author calls STRIDE "a solid foundation" for agents, but one that "doesn't address the unique challenges of AI agents", such as the risks that come from their unpredictable decision-making.

So add a seventh row: what if the agent is simply wrong? No access rule fixes it, because the access was correct. The control is an independent evaluation of what the agent produced, against what it was supposed to do, before anyone relies on it.

Did we do a good enough job? Keep the threat model current

An agent's threat model expires faster than most. A new tool adds a boundary, a new model changes behaviour, and a widened scope changes what an elevation of privilege can reach. Re-run the walkthrough whenever any of those change, and turn each threat you found into a test that runs on every release, the way red-team findings should become release gates.

Where Syntrox fits

A judge that evaluates an agent's output sits in the threat model too, so it has to resist the same tampering. Nucleus treats any text inside the material it evaluates as evidence, never as instructions, and an instruction addressed to the evaluator counts against the output; its injection-resistance probes run as a release gate with explicit thresholds. It is small enough to run on hardware inside your own network, so the data a check needs does not cross a new trust boundary, and Syntrox records every verdict outside the agent with the judge version that produced it.