Security Pass

Context

An MCP server is a new attack surface with unique properties — tool output feeds back into the LLM's context, scopes gate what the model can do on the user's behalf, and per-request state must stay tenant-scoped. This skill walks a server through eight axes shaped around what the server builder actually controls. Framework-level concerns (transport, JSON-RPC parsing, auto-correlation, error classification) are out of scope — mcp-ts-core handles those.

Read the code. Don't trust patterns from memory.

When to Use

• Before a release
• After adding or modifying a batch of handlers or services
• Periodically (quarterly-ish)
• User asks for a "security review", "audit", "hardening pass", or similar

Inputs

Gather before starting. Ask if unclear:

1. Scope — whole server, specific module, or recent diff?
2. Known concerns — anything the user already suspects?
3. Deployment context — multi-tenant? public network? auth mode? (stdio / local-http / public-http behave differently)
4. Severity floor — report all findings, or skip medium/low?

Steps

1. Build the map

Surface what you're auditing before diving in. Paths below assume the mcp-ts-core layout — adjust to your repo.

find src/mcp-server/tools/definitions -name "*.tool.ts" | sort
find src/mcp-server/resources/definitions -name "*.resource.ts" 2>/dev/null | sort
find src/mcp-server/prompts/definitions -name "*.prompt.ts" 2>/dev/null | sort
find src/services -maxdepth 1 -mindepth 1 -type d | sort

Note: tool / resource / prompt counts, auth mode, storage provider, upstream APIs, which tools have destructiveHint, which handlers request a consent round via ctx.requestInput, which services hold module-scope state, whether the server reads roots.

If transport is streamable HTTP or SSE, also capture:

• Bind address (127.0.0.1 for local, or 0.0.0.0 / public interface?)
• Origin allowlist (DNS rebinding mitigation) — configured, or wildcard / missing?
• Session ID source (framework CSPRNG, or builder-supplied?) and binding to auth identity
• Any unauthenticated routes (/healthz, /sse, metadata endpoints) — do they leak tool lists or tenant hints?
• MCP Authorization spec: if implemented, PKCE enforced, token audience (aud) checked, resource indicators used

If CANVAS_PROVIDER_TYPE=duckdb is set, also capture:

• Auth mode — canvas in MCP_AUTH_MODE=none collapses the composite (tenantId, canvasId) scope to ('default', canvasId), where the ID is the only differentiator
• CANVAS_MAX_CANVASES_PER_TENANT, CANVAS_TTL_MS, CANVAS_ABSOLUTE_CAP_MS, CANVAS_EXPORT_PATH values
• Whether external rate limiting (CDN, reverse proxy, WAF) fronts the deployment — required to keep the ~10¹⁸ canvasId keyspace operationally infeasible to brute-force

Use TaskCreate — one task per axis. Mark complete as you go.

Run fuzzTool in parallel. @cyanheads/mcp-ts-core/testing/fuzz catches crashes, memory leaks, and prototype pollution automatically on each tool — start it now so results are ready when you reach Axis 5.

2. Walk the eight axes

Axis 1 — LLM-facing surfaces as injection vector

Anything the server sends to the client that reaches the LLM's context is a potential injection surface: tool output, resource content, prompt text, and the metadata the LLM reads to decide what to call. Relayed upstream content (tickets, scraped text, emails, DB rows) can carry adversarial instructions even when your code is honest.

Look in:

• Every *.tool.ts — output schema + format()
• Every *.resource.ts — content returned from resources/read
• Every *.prompt.ts — templated message content
• Every definition file — description, title, annotations, and inputSchema field descriptions (templated from untrusted data?)

Check:

• Handlers that return raw upstream text / DB rows without structural framing?
• Does format() wrap untrusted content in delimiters (blockquote, fenced code, <data> tags)?
• Output schema distinguishes "data" fields from free-form text?
• Resource content (resources/read) framed the same way tool output is?
• Prompt templates interpolate untrusted data without escaping — treating tenant-controlled strings as trusted instructions?
• Tool / resource / prompt descriptions templated from runtime data? Static strings are safer; templated descriptions enable "tool poisoning" (adversarial metadata steering the LLM toward a dangerous tool).
• Descriptions mutated mid-session? Rug-pull surface: client approved the v1 description, server now advertises v2 behavior.

Smell: return { body: await fetch(url).then(r => r.text()) } rendered directly in format(). Or: description: \Look up ${tenant.customLabel}`wherecustomLabel` is tenant-supplied.

Axis 2 — Scope granularity

Every auth: [...] entry is a blast-radius dial.

Look in: every *.tool.ts — auth: array.

grep -rn "auth: \[" src/mcp-server/tools/definitions/

Check:

• Tools with ['admin'], ['*'], or []?
• A single scope covering two capabilities that should be separated (read vs write)?
• Read-only tools never require write scopes?
• MCP_AUTH_DISABLE_SCOPE_CHECKS=true set in production? When on, both withRequiredScopes and checkScopes early-return — every authenticated user gets every tool, and runtime tenant patterns like team:${input.teamId}:write no longer guard. Acceptable only when paired with a real server-side ACL (path filter, allowlist, upstream API enforcement).

Smell: every tool shares the same scope string. Or: MCP_AUTH_DISABLE_SCOPE_CHECKS=true set without a documented compensating ACL — confirm the deployment relies on a meaningful access control layer below the framework before approving.

Axis 3 — Destructive ops without a consent round

ctx.requestInput moves consent off the LLM and onto the user: the handler returns an input_required result and only runs the side effect once it is re-entered and redeems the record it stored when it asked. An accepted response on ctx.inputs alone proves nothing — a client can send one on a call nothing prompted for. Destructive tools without that round trust the LLM not to be tricked.

Look in: handlers with destructiveHint: true or side-effecting verbs in names (delete_*, send_*, pay_*, publish_*, drop_*).

grep -rn "destructiveHint" src/mcp-server/tools/definitions/
grep -rnE "ctx\.requestInput|ctx\.inputs" src/mcp-server/tools/definitions/

Check:

• Each destructive handler reads ctx.inputs for the confirmation and returns ctx.requestInput(...) before the side effect?
• The confirmation response is validated — ctx.inputs.accepted(key, Schema) with a schema, not the bare overload. The SDK never re-validates a response against the schema its request advertised, and the payload is LLM-mediated: "user confirmed" does not mean "user authored these exact fields."
• A declined or cancelled response is terminal — checked via ctx.inputs.view(key) and thrown on, never re-asked (a re-ask loops until the round budget runs out) and never treated as consent.
• Consent is scoped to the specific target (e.g., record ID rendered in the message), not a generic "proceed?"
• Where requestState influences authorization, resource access, or which target gets mutated, MCP_REQUEST_STATE_KEY is set (≥ 32 bytes, the same on every instance a retry can reach). The framework then seals the string a handler returns and the SDK rejects any other state before the handler runs; unset, the state round-trips through the client and comes back attacker-controlled.
• A consent gate's state is server-issued and single-use. The framework drops answers the client's declared capabilities do not cover, kind and mode, on both eras, so a client without elicitation.form — a URL-only client included — cannot pre-answer a form gate. A client that declared it still can: it may send inputResponses — and a requestState of its own, or one it was issued earlier — on the very first call, so a handler that only compares client-carried state against a fresh resolution deletes on an answer nobody was shown. A sealed state closes forgery but not replay within its 900 s lifetime. Keep what the prompt confirmed in a ctx.state record keyed by a random id — the operation (tool name, or the resource URI read), the caller (ctx.auth clientId and sub), the target, and a content hash so a same-path swap is caught — send only the id, redeem it before anything else in the handler, and ask again on an unknown, used, or expired id or on any field that differs from this call. Without the operation, an id minted by another gated tool or resource confirms this one; without the caller, another user in the same tenant redeems an id they were handed while MCP_REQUEST_STATE_KEY is unset. The record's storage is shared by every instance a 2026-07-28 retry can reach — filesystem, supabase, or cloudflare-d1, never cloudflare-kv, whose eventual consistency widens the race below.
• Redeeming is not atomic. Read-then-delete stops a sequential replay, but concurrent retries carrying one id can each read the record before any delete lands and each run the action. Until ctx.state gains an atomic take (#593), an action that must not repeat — a payment, a send, a publish — is idempotent per record (the record id as the upstream idempotency key), or the risk is accepted knowingly and recorded in the findings.
• The weak point is answerability, not availability. ctx.requestInput is present on every transport and both protocol eras — the 2025-era shim issues the real elicitation/create round trip, the 2026-07-28 client fulfils the embedded request directly. A client that never retries simply leaves the destructive step un-run, which fails safe. Keep destructiveHint: true so client-side approval flows still surface the risk, and do not accept "proceed anyway when the round is unavailable" as a fallback. On a 2025-era connection whose client lacks the capability, ctx.requestInput throws client_capability_missing inside the handler; catching that to run the side effect is exactly this bypass — let it propagate. So is skipping the prompt because ctx.clientCapabilities lacks elicitation: that property decides whether to ask for optional context, never whether consent is needed.

Smell: destructiveHint: true file with no ctx.requestInput in it. Or ctx.inputs.accepted('confirm') with no schema argument — the content could be anything. Or a gate that proceeds on ctx.inputs without redeeming a ctx.state record, whose record omits the operation or the caller, or that compares a target carried in requestState. Or a non-repeatable action behind a gate with no idempotency key. Or a handler that re-issues the same request after a decline.

Axis 4 — Upstream auth shape

What credentials the server holds, and the blast radius if one leaks.

Look in: src/services/*, src/config/server-config.ts.

Check:

• Each upstream API key scoped to minimum required? (No admin keys for read workflows.)
• Services re-mint downstream tokens with correct aud, or passthrough the caller's?
• Server holds OAuth for N services × M tenants — what does one-tenant compromise expose?
• Per-tenant rate limits on upstream calls?

Smell: one global API_KEY used across all tenants + retry loop with no upper bound.

Axis 5 — Input sinks

LLM-supplied inputs feel internal but aren't. Classic sinks apply, amplified. Sampling responses and roots-derived paths are MCP-specific sinks that look internal but carry LLM/client trust.

Look in: all handlers.

# URL sinks — SSRF
grep -rn "z.string().url()" src/

# Path sinks — traversal
grep -rnE "readFile|writeFile|readdirSync|createReadStream|statSync" src/

# Shell sinks — command injection
grep -rnE "\b(exec|spawn|execSync|spawnSync)\b" src/

# Merges — prototype pollution
grep -rnE "Object\.assign\b|structuredClone" src/

# Lookups — prototype chain read through an object literal
grep -rnE "\[[a-zA-Z_$][a-zA-Z0-9_$.]*\] *\?\? |\[[a-zA-Z_$][a-zA-Z0-9_$.]*\] *\|\| " src/

# Roots — client-shared filesystem
grep -rnE "roots/list|ctx\.roots" src/

# Schema laxity — fields sneaking past validation
grep -rnE "\.passthrough\(\)|\.loose\(\)|looseObject\(|\.catchall\(" src/mcp-server/

Check:

• URL-taking tools block private IPs, file://, ftp://, localhost, DNS rebind?
• Path-taking tools canonicalize (path.resolve + assert startsWith(root + sep))?
• Upstream URL paths built from caller segments refuse . and ..? encodeURIComponent leaves dots untouched, and the URL parser resolves dot segments, so a file key or archive-member input of ../../me retargets the request (credentials attached) at another endpoint on the same host. Redirects on authenticated requests should follow only within the upstream origin.
• Caller text spliced into an upstream query language next to server-built filter clauses stays well-formed? Some search backends fall back to plain-text matching when they can't parse a query. An unclosed quote or a trailing \ from the caller then silently drops every filter clause while the tool still reports them as applied.
• Roots-derived paths: resolved result stays within one declared root (iterate and assert), not assumed-safe because "the client said so"?
• Shell-using tools use an allowlist (never string-concat)?
• Regex / glob / filter inputs bounded (length cap, complexity limits, execution timeout) — ReDoS-safe?
• The server's own patterns are linear-time on hostile input? Every .regex() in an input schema, and every regex a handler or normalizer runs over caller text, sees attacker-length strings before any length cap applies. Loose "raw" patterns that admit un-normalized input (\s* runs, optional quotes and separators around a repeated group) are the usual source of polynomial backtracking. Time each one against a long adversarial string (thousands of spaces, then a character that forces failure). Milliseconds is fine; seconds is a finding.
• User-JSON merges reject __proto__, constructor, prototype keys?
• Lookup tables keyed by input-derived text are a Map, not an object literal? The read direction of the same defect: TABLE[key] ?? fallback walks the prototype chain, so a key of constructor (or toString, valueOf) returns a function the ?? does not catch — it is not nullish — and string coercion then emits function Object() { [native code] } into the output. Lowercasing the key masks the camelCase members and leaves constructor reachable, so it is not a fix. A Map has no prototype chain; Object.create(null) works too.
• Input schemas .strict() — unknown fields rejected, not silently passed to downstream code that destructures with ...rest?
• Output schemas without .passthrough() / .loose() / .catchall() — no accidental exfiltration of fields your schema didn't declare? Smell: z.string().url() with no allowlist; readFile(input.path) with no canonicalization.