GHSA-r5pq-6chh-j3xp: CSRF

Summary

An unauthenticated POST to any OpenAPI-validated endpoint, including the anonymous POST /edge/validate and POST /edge/issue-token, crashes the entire Unleash server with one request body of deeply-nested JSON.

When request-body validation fails, Unleash builds the error message by calling JSON.stringify on the raw offending value taken from the request body. A value nested a few thousand levels deep makes JSON.stringify recurse past the V8 call-stack limit and throw RangeError: Maximum call stack size exceeded.

The throw is synchronous inside an Express error-handling middleware that has no try/catch, and the process registers no uncaughtException handler (only unhandledRejection). Node terminates the process with exit(1).

The body parser sets no JSON nesting-depth limit, the payload (about 10 KB at depth 5000) stays far under the 100 KB body-size limit, and the same error formatter serves every API route, so the trigger is route-independent.

Result: a remote attacker with no account, token, or cookie takes the whole server offline with a single 10 KB request, and keeps it offline by replaying it.

Affected

Unleash OSS server, confirmed live-exploitable on v8.0.0 (unleashorg/unleash-server:8.0.0). Vulnerable code is the shared request-validation error path (src/lib/error/bad-data-error.ts genericErrorMessage), present on current main. Reachable on a stock install: no feature flag, no setting, no authentication, no CSRF token, and no cookie required. The 100 KB request-body size limit does not mitigate it; the crashing payload is about 10 KB.

Root cause

On validation failure the error formatter serializes the raw offending request value with an unguarded const youSent = JSON.stringify(propertyValue) (src/lib/error/bad-data-error.ts:75), where propertyValue is read straight from the request body via lodash.get (bad-data-error.ts:123). A deeply-nested array or object makes JSON.stringify recurse once per level and throw RangeError: Maximum call stack size exceeded. fromOpenApiValidationErrors (bad-data-error.ts:158) runs from the Express error middleware openAPIValidationMiddleware (src/lib/routes/controller.ts:68), which calls it with no try/catch (controller.ts:70), while the controller's own try/catch (controller.ts:103) wraps only the route handler. The process registers only process.on('unhandledRejection') (src/lib/server-impl.ts:274) and no uncaughtException handler, so the synchronous throw terminates Node with exit(1). The same unguarded pattern exists at fromJoiError (bad-data-error.ts:175) and at handleErrors (src/lib/routes/util.ts:42), whose own comment records that JSON.stringify(finalError.details) "also hangs" (util.ts:63).

Reproduction

unleashorg/unleash-server:8.0.0 started from the repository docker-compose.yml, default config, no credentials, server at http://localhost:4242.

1. A shallow body returns a normal 400 and the server stays up.

curl -s -X POST http://localhost:4242/edge/validate \ -H 'Content-Type: application/json' -d '{"tokens":[[]]}' HTTP 400 BadDataError: "The /body/tokens/0 property must be string. You sent []."

2. A single anonymous body nested 5000 levels deep (10013 bytes) crashes the process.

curl -s -o /dev/null -w '%{httpcode}\n' -X POST http://localhost:4242/edge/validate \ -H 'Content-Type: application/json' \ --data-binary "$(python3 -c 'd=5000;print("{\"tokens\":["+"["d+"]"d+"]}")')" 000 (connection dropped mid-response; Node exited)

3. The server is gone and does not recover on its own.

curl -s -o /dev/null -w '%{httpcode}\n' http://localhost:4242/health 000 docker inspect repo-web-1 --format '{{.State.Status}} restarts={{.RestartCount}}' exited restarts=0

Live-verified: one anonymous 10013-byte request to /edge/validate drove /health from 200 to 000; the container logs showed the RangeError stack through bad-data-error and openAPIValidationMiddleware, and the process did not restart. The crash threshold is about 4000 to 4500 nesting levels; depth 5000 crashes reliably. The same crash reproduces on POST /edge/issue-token with a nested object, confirming route and shape independence.

Impact

- Single 10 KB request with no account, token, or cookie takes the entire server offline. - Replaying the request (kilobits per restart interval) sustains a near-total outage; with no restart policy the first request is a permanent kill. - Feature-flag evaluation sits on the request path of dependent applications, so the outage propagates to them. - Every OpenAPI-validated endpoint is a trigger, including anonymous ones.

Credit

Jan Kahmen, turingpoint (jan@turingpoint.de)