GHSA-GCGF-FH7C-8GF2
Vulnerability from github – Published: 2026-10-05 17:32 – Updated: 2026-10-05 17:32Summary
ZodSmartCoercionPlugin and experimental_ZodSmartCoercionPlugin mishandle object keys that name Object.prototype members. Both coerce request input before validation, so any client that can reach a procedure whose input schema contains an object or a record can:
- replace the prototype of the coerced input object, and
- make a request fail with an unhandled
TypeErrorby sending a key such asconstructor.
This is the same class of bug as GHSA-4h5r-cv8j-4456 in @orpc/json-schema, in a different package.
This is not global prototype pollution
The global Object.prototype is never modified. No shared or global state is written, so other objects in the process, other requests, and other users are not affected.
What changes is the prototype of one object: the value the coercer returns for the request that carried the payload. That object is passed on to validation and to the procedure handler, then discarded when the request ends. An ordinary {} created anywhere else in the application is unaffected.
Details
1. __proto__ replaces the prototype of the coerced value
Coerced properties were collected into a plain {}, so newObj['__proto__'] = value triggers the inherited __proto__ setter instead of creating an own property. Verified on 1.14.8 against a record schema, in both plugins:
schema: z.record(z.string())
input: {"a":"1","__proto__":{"polluted":true}}
Object.keys(result) => ["a"] // the __proto__ entry is dropped from the output
result.polluted => true // the attacker object is now this object's prototype
({}).polluted => undefined // Object.prototype is untouched
The returned value inherits attacker controlled properties, and any consumer reading them through the prototype chain, such as a handler checking input.isAdmin, an Object.assign copy, or a config lookup, sees data the client never legitimately supplied. How far that goes depends on what the application does with the coerced input. There is no generic escalation path.
This affects the record branch of both plugins, and the object branch of both plugins for any key that does not first trigger the error below.
2. Sub-schemas resolved through the prototype chain
The object branch looked up sub-schemas with shape[key], which returns inherited members for keys such as constructor, toString and __proto__. Those values, for example the Object constructor, were then dereferenced as Zod schemas:
schema: z.object({ a: z.number() })
input: { a: '123', constructor: '456' }
zod 3 => TypeError: Cannot read properties of undefined (reading 'Symbol(ORPC_CUSTOM_ZOD_DEF)')
zod 4 => TypeError: Cannot read properties of undefined (reading 'def')
Coercion throws before the assignment, so for object schemas these keys break the request rather than injecting a prototype. Any client can trigger this on its own requests. It does not affect other clients or the process as a whole.
Impact
Any oRPC server that installs ZodSmartCoercionPlugin or experimental_ZodSmartCoercionPlugin from @orpc/zod and has at least one procedure whose input schema contains an object or a record. No authentication or user interaction is required beyond what the procedure itself demands.
The effect is scoped to the request that sent the payload. Because no global state is written, one request cannot influence another, and restarting the process is not needed to recover.
Patch
Fixed in 1.14.10.
In packages/zod/src/coercer.ts and packages/zod/src/zod4/coercer.ts:
- collect coerced object and record properties into
NullProtoObj, the null prototype helper already used for untrusted keys inbracket-notation.tsandrpc-matcher.ts, so__proto__stays an ordinary own property and is preserved in the output - guard the shape lookup with
Object.hasOwn, so unknown keys fall through tocatchallas they already do for every other name
PR: middleapi/orpc#1727
Workarounds
Reject or strip request payloads containing __proto__, constructor or prototype keys before they reach the handler, or remove the plugin if automatic coercion is not required.
Note that SmartCoercionPlugin from @orpc/json-schema, the recommended replacement for these plugins, was affected by the same class of issue (GHSA-4h5r-cv8j-4456) and is fixed in the same release.
Credit
Reported and fixed internally.
{
"affected": [
{
"package": {
"ecosystem": "npm",
"name": "@orpc/zod"
},
"ranges": [
{
"events": [
{
"introduced": "0"
},
{
"fixed": "1.14.10"
}
],
"type": "ECOSYSTEM"
}
]
}
],
"aliases": [
"CVE-2026-103918"
],
"database_specific": {
"cwe_ids": [
"CWE-1321",
"CWE-915"
],
"github_reviewed": true,
"github_reviewed_at": "2026-10-05T17:32:32Z",
"nvd_published_at": "2026-10-02T20:17:00Z",
"severity": "MODERATE"
},
"details": "## Summary\n\n`ZodSmartCoercionPlugin` and `experimental_ZodSmartCoercionPlugin` mishandle object keys that name `Object.prototype` members. Both coerce request input before validation, so any client that can reach a procedure whose input schema contains an object or a record can:\n\n1. replace the prototype of the coerced input object, and\n2. make a request fail with an unhandled `TypeError` by sending a key such as `constructor`.\n\nThis is the same class of bug as GHSA-4h5r-cv8j-4456 in `@orpc/json-schema`, in a different package.\n\n## This is not global prototype pollution\n\n**The global `Object.prototype` is never modified.** No shared or global state is written, so other objects in the process, other requests, and other users are not affected.\n\nWhat changes is the prototype of one object: the value the coercer returns for the request that carried the payload. That object is passed on to validation and to the procedure handler, then discarded when the request ends. An ordinary `{}` created anywhere else in the application is unaffected.\n\n## Details\n\n### 1. `__proto__` replaces the prototype of the coerced value\n\nCoerced properties were collected into a plain `{}`, so `newObj[\u0027__proto__\u0027] = value` triggers the inherited `__proto__` setter instead of creating an own property. Verified on 1.14.8 against a record schema, in both plugins:\n\n```\nschema: z.record(z.string())\ninput: {\"a\":\"1\",\"__proto__\":{\"polluted\":true}}\n\nObject.keys(result) =\u003e [\"a\"] // the __proto__ entry is dropped from the output\nresult.polluted =\u003e true // the attacker object is now this object\u0027s prototype\n({}).polluted =\u003e undefined // Object.prototype is untouched\n```\n\nThe returned value inherits attacker controlled properties, and any consumer reading them through the prototype chain, such as a handler checking `input.isAdmin`, an `Object.assign` copy, or a config lookup, sees data the client never legitimately supplied. How far that goes depends on what the application does with the coerced input. There is no generic escalation path.\n\nThis affects the `record` branch of both plugins, and the `object` branch of both plugins for any key that does not first trigger the error below.\n\n### 2. Sub-schemas resolved through the prototype chain\n\nThe `object` branch looked up sub-schemas with `shape[key]`, which returns inherited members for keys such as `constructor`, `toString` and `__proto__`. Those values, for example the `Object` constructor, were then dereferenced as Zod schemas:\n\n```\nschema: z.object({ a: z.number() })\ninput: { a: \u0027123\u0027, constructor: \u0027456\u0027 }\n\nzod 3 =\u003e TypeError: Cannot read properties of undefined (reading \u0027Symbol(ORPC_CUSTOM_ZOD_DEF)\u0027)\nzod 4 =\u003e TypeError: Cannot read properties of undefined (reading \u0027def\u0027)\n```\n\nCoercion throws before the assignment, so for object schemas these keys break the request rather than injecting a prototype. Any client can trigger this on its own requests. It does not affect other clients or the process as a whole.\n\n## Impact\n\nAny oRPC server that installs `ZodSmartCoercionPlugin` or `experimental_ZodSmartCoercionPlugin` from `@orpc/zod` and has at least one procedure whose input schema contains an object or a record. No authentication or user interaction is required beyond what the procedure itself demands.\n\nThe effect is scoped to the request that sent the payload. Because no global state is written, one request cannot influence another, and restarting the process is not needed to recover.\n\n## Patch\n\nFixed in 1.14.10.\n\nIn `packages/zod/src/coercer.ts` and `packages/zod/src/zod4/coercer.ts`:\n\n- collect coerced object and record properties into `NullProtoObj`, the null prototype helper already used for untrusted keys in `bracket-notation.ts` and `rpc-matcher.ts`, so `__proto__` stays an ordinary own property and is preserved in the output\n- guard the shape lookup with `Object.hasOwn`, so unknown keys fall through to `catchall` as they already do for every other name\n\nPR: middleapi/orpc#1727\n\n## Workarounds\n\nReject or strip request payloads containing `__proto__`, `constructor` or `prototype` keys before they reach the handler, or remove the plugin if automatic coercion is not required.\n\nNote that `SmartCoercionPlugin` from `@orpc/json-schema`, the recommended replacement for these plugins, was affected by the same class of issue (GHSA-4h5r-cv8j-4456) and is fixed in the same release.\n\n## Credit\n\nReported and fixed internally.",
"id": "GHSA-gcgf-fh7c-8gf2",
"modified": "2026-10-05T17:32:32Z",
"published": "2026-10-05T17:32:32Z",
"references": [
{
"type": "WEB",
"url": "https://github.com/middleapi/orpc/security/advisories/GHSA-gcgf-fh7c-8gf2"
},
{
"type": "ADVISORY",
"url": "https://nvd.nist.gov/vuln/detail/CVE-2026-103918"
},
{
"type": "WEB",
"url": "https://github.com/middleapi/orpc/pull/1727"
},
{
"type": "WEB",
"url": "https://github.com/middleapi/orpc/commit/26314dfb443237c495116c5794d3d30a7a22c570"
},
{
"type": "PACKAGE",
"url": "https://github.com/middleapi/orpc"
},
{
"type": "WEB",
"url": "https://github.com/middleapi/orpc/releases/tag/v1.14.10"
}
],
"schema_version": "1.4.0",
"severity": [
{
"score": "CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:L/I:L/A:N",
"type": "CVSS_V3"
}
],
"summary": "@orpc/zod: Prototype injection in smart coercion"
}
Sightings
| Author | Source | Type | Date | Other |
|---|
Nomenclature
- Seen: The vulnerability was mentioned, discussed, or observed by the user.
- Confirmed: The vulnerability has been validated from an analyst's perspective.
- Published Proof of Concept: A public proof of concept is available for this vulnerability.
- Exploited: The vulnerability was observed as exploited by the user who reported the sighting.
- Patched: The vulnerability was observed as successfully patched by the user who reported the sighting.
- Not exploited: The vulnerability was not observed as exploited by the user who reported the sighting.
- Not confirmed: The user expressed doubt about the validity of the vulnerability.
- Not patched: The vulnerability was not observed as successfully patched by the user who reported the sighting.
The approach is described in our paper Mapping CVEs to MITRE ATT&CK Techniques: A Curated Gold-Set Classifier and the Limits of LLM-Assisted Label Expansion.
Browse all ATT&CK techniques and the vulnerabilities related to each.
Related by attack behaviour
Vulnerabilities whose description is nearest to this one in the vector space of the CIRCL/vulnerability-attack-technique-biencoder model. This is a similarity search over the bi-encoder space (plain cosine), not a classification, and it has no measured accuracy.