MCP / Agent Server Security Checklist
Audit-ready security checklists and a STRIDE threat model for remote MCP servers and agent-facing APIs: auth, input validation, rate limiting, secrets, payments, and data-exposure controls.
File Structure 7 files
Documentation Preview README excerpt
MCP / Agent Server Security Checklist & Threat Model
A practical, audit-ready security checklist and threat model for anyone running a
remote MCP server or an agent-facing HTTP API. Agent traffic is a new
attack surface: the "user" may be an autonomous model steered by an untrusted
prompt. This pack is the checklist used to harden a live, payment-handling MCP
server before it went public.
What's inside
checklists/pre-launch.md— a gating checklist to run before you expose an
endpoint to agents (auth, input validation, rate limiting, secrets, CORS).
checklists/payments.md— extra controls when an agent can trigger money
movement (price-tampering, idempotency, webhook verification, no autonomous
charging).
checklists/data-exposure.md— preventing leakage of secrets, signed URLs,
storage paths, PII, and other-customer data through tools/previews.
THREAT-MODEL.md— the STRIDE-style threat model with concrete mitigations,
mapped to the checklists.
Who this is for
- Engineers shipping a remote MCP server or agent API.
- Security reviewers auditing one.
- Anyone letting an agent take an action that costs money or touches user data.
The one-paragraph version
Treat every agent request as potentially prompt-injected and hostile. Validate
and constrain all inputs at the boundary. Resolve anything sensitive (price, ids,
file paths) server-side — never from the caller. Rate-limit and log without
recording PII. Keep secrets and signed URLs out of every tool response. Require
buyer authorization for payments; never charge autonomously. Verify webhooks with
the signing-secret flow. Scope every "get my stuff" tool to a proven owner.
Requirements
None — this is documentation you apply to your own stack. No dependencies.
File tree
checklists/
pre-launch.md # gate before exposing to agents
payments.md # money-movement controls
data-exposure.md # secret / PII / path leakage prevention
THREAT-MODEL.md # STRIDE threats + mitigations
License
MIT — copy these into your own runbooks and repos.
Content Sample THREAT-MODEL.md
Threat Model — Remote MCP / Agent-Facing API
A STRIDE-style threat model for an agent-facing service. The novel element vs a
normal API: the "user" is often an LLM steered by untrusted input (web pages,
documents, other tools). Treat every request as potentially prompt-injected.
Assets
- Payment capability (ability to create charges).
- Customer entitlements and downloadable artifacts.
- Secrets (processor keys, database keys, signing secrets).
- Catalogue integrity (price, sellability, verification status).
Trust boundaries
- Agent client → your endpoint (fully untrusted; may be adversarial).
- Your endpoint → payment processor (trusted, key-authenticated).
- Payment processor → your webhook (trusted only after signature verification).
STRIDE
Spoofing
- Threat: caller claims to be a user/owner it is not.
- Mitigation: ownership via unguessable capability tokens or verified auth
subject; never trust a client-supplied user id. (data-exposure.md, pre-launch.md)
Tampering
- Threat: caller alters price/amount/currency/ids to underpay or buy blocked
items; alters redirect to phish.
- Mitigation: resolve all money + ids server-side; reject client money fields;
allowlist redirects; DB constraint blocking sale of unaudited/deprecated items.
(payments.md)
Repudiation
- Threat: disputed or unattributable actions.
- Mitigation: structured, PII-free audit logs; processor metadata (order id,
agent id) on every payment. (pre-launch.md)
Information disclosure
- Threat: leakage of secrets, signed URLs, storage paths, PII, or other
customers' data via tools/previews/errors.
- Mitigation: never return secrets/signed URLs; preview sanitizer; owner-scoped
reads; redacting logs; safe error messages. (data-exposure.md)
Denial of service
- Threat: high-volume agent traffic exhausts resources or budget (e.g. mass
checkout-session creation).
- Mitigation: per-IP/key rate limits, tighter on mutations; cheap read paths
... and much more in the full download.