high

CVE-2026-73654

npm · @trigger.dev/core

Summary

Trigger.dev: Prototype pollution via run metadata operations → process-wide cross-tenant DoS

Severity
high
CVSS
8.5
EPSS
0.6% (p49)
CWE
CWE-1321
Also known as
GHSA-p28v-f755-9qrg
Published
2026-08-13
Updated
2026-08-13

Advisory details

Summary

The run-metadata update endpoint PUT /api/v1/runs/:runId/metadata applies client-supplied "operations" by passing the attacker-controlled operation.key straight into new JSONHeroPath(operation.key).set(newMetadata, value) (packages/core/src/v3/runMetadata/operations.ts:22-23), with no prototype-pollution guard (@jsonhero/path@^1.0.21 does not reject __proto__/constructor/prototype).

A request with key: "$.__proto__.polluted" sets Object.prototype.polluted in the webapp process. Because every plain object then inherits that property, it corrupts unrelated code process-wide and across tenants — including Prisma query building and the Prometheus metrics client — causing query failures, broken authentication for other tenants' workers, and an uncaughtException (denial of service). Only a normal, low-privilege environment API key is required (one request).

Severity

A single request from any holder of a normal environment API key contaminates Object.prototype in the shared webapp process, breaking other tenants' workers (scope change) and degrading/ crashing the process (high availability impact). Prototype pollution is also a primitive for further gadget chains (integrity/confidentiality).

Affected versions

Root cause

packages/core/src/v3/runMetadata/operations.ts — applyMetadataOperations() builds a path from the untrusted operation.key and writes to it, for every operation type (set, append, increment, …):

const path = new JSONHeroPath(operation.key);   // operation.key fully attacker-controlled
path.set(newMetadata, operation.value);         // no __proto__/constructor/prototype rejection

The request schema (UpdateMetadataRequestBody) types key as a plain string with no validation, and @jsonhero/path@1.0.21 walks __proto__ as an ordinary segment → the assignment lands on Object.prototype.

Proof of Concept

Self-host ghcr.io/triggerdotdev/trigger.dev:v4-beta. Authenticated with a normal environment API key (tr_dev_…) and any run id of that environment.

curl -X PUT "http://localhost:8030/api/v1/runs/run_cmqr2bsyo00013js2twwhdsfu/metadata" \
  -H "Authorization: Bearer tr_dev_<env_key>" -H "Content-Type: application/json" \
  --data '{"operations":[{"type":"set","key":"$.__proto__.polluted","value":"PWNED"}]}'

Result — Object.prototype.polluted = "PWNED" process-wide. Observed in the webapp logs:

  1. The request's own query is corrupted — polluted:"PWNED" injected into every object Prisma enumerates:
    prisma.taskRun.updateMany({ where:{ id:"…", metadataVersion:2, polluted:"PWNED" },
      data:{ …, metadataVersion:{ increment:1, polluted:"PWNED" }, polluted:"PWNED" }, polluted:"PWNED" })
    -> Unknown argument `polluted`
    
  2. Cross-tenant authentication break — the next request from a different client (a worker's POST /engine/v1/dev/dequeue) fails inside findEnvironmentByApiKey:
    prisma.runtimeEnvironment.findFirst({ where:{ apiKey:"…", polluted:"PWNED" },
      include:{ project:true, …, polluted:"PWNED" } })  -> PrismaClientValidationError
    
    i.e. one tenant's request breaks authentication for other tenants' workers → their jobs stop being dequeued/processed.
  3. Denial of service — full process crash. The uncaughtException in prom-client (Error: Added label "polluted" is not included in initial labelset: [ 'kind' ]) crashes the webapp process. Demonstrated with a second tenant: Tenant B (a different org/env, with its own API key) had a working request (POST /engine/v1/dev/dequeue → HTTP 400, auth OK) before the attack; immediately after Tenant A's single attack request, B's request returned HTTP 000 (no response) — the whole multi-tenant webapp was down. Logs show the crash at 20:53:41 and the process auto-restarting ~3s later (FairQueue/ScheduleEngine started). Repeating the attack in a loop yields a crash-loop = sustained DoS for all tenants.

(Note: the metadata endpoint itself swallows the Prisma failure with ignoreError:true and still returns HTTP 200 — the damage is the process-wide contamination observed in the logs, not the endpoint's status code.)

Impact

A low-privilege caller (one normal environment API key, one request) pollutes Object.prototype in the shared multi-tenant webapp process, causing:

Suggested remediation

  1. Reject dangerous path segments in operation.key before building the path — block __proto__, constructor, prototype (and validate the $.-rooted JSONHero path shape).
  2. Build metadata on a null-prototype object (Object.create(null)) and/or use a pollution-safe setter, so __proto__ cannot reach Object.prototype.
  3. Upgrade/replace @jsonhero/path for a version that is prototype-pollution safe, or wrap its .set() with a guard.

References

Related advisories

Is your project exposed to this? Stateward checks every dependency on every pull request and flags it only if your code actually reaches it.

Check my repo

Summarize with AI

ChatGPTClaudePerplexity

Sources: CISA KEV (public domain), OSV.dev & GitHub Advisory Database (CC-BY-4.0), FIRST EPSS, NVD/CWE (public domain). Served live from the Stateward advisory database.