GHSA-2CP2-2R3C-7P7R
Vulnerability from github – Published: 2026-08-21 20:57 – Updated: 2026-08-21 20:57Summary
hydra.utils.instantiate() resolves and calls Python objects from config. If an
application passes untrusted config to instantiate(), an attacker who controls
_target_ and its arguments can cause arbitrary code execution in the consuming
process.
Hydra is not a network service. Exploitation requires a consuming application,
library, or user workflow to load attacker-controlled config, CLI overrides, or
model metadata and pass it to hydra.utils.instantiate().
Details
Hydra's instantiate API is designed to construct objects and call functions from configuration. For example:
component:
_target_: package.module.Class
arg: value
When this config is passed to hydra.utils.instantiate(), Hydra resolves
_target_ and calls it with the provided arguments.
This is intended for trusted application configuration. However, if untrusted
input controls _target_, the config becomes a callable-selection mechanism. A
malicious config can select a callable capable of executing code or commands and
provide attacker-controlled arguments.
This issue is the same general class of problem discussed by Unit 42 for downstream AI/ML libraries such as NVIDIA NeMo, where untrusted model metadata was passed into Hydra instantiate:
https://unit42.paloaltonetworks.com/rce-vulnerabilities-in-ai-python-libraries/
Hydra 1.3.4 includes a blacklist for some dangerous _target_ values. That
blacklist is defense-in-depth and is not a complete security boundary. The
blacklist is not present in the released hydra-core 1.3.3 package, so this
issue should not be described as a bypass of a released 1.3.3 blacklist.
Impact
A successful attack can execute code in the process that calls
hydra.utils.instantiate(). The impact is limited to the privileges and
environment of that process.
Potential impact includes:
- Reading files, credentials, environment variables, or data accessible to the process
- Modifying files, outputs, checkpoints, or application state writable by the process
- Terminating or disrupting the process
Affected Usage
Applications and libraries are affected when they pass untrusted or semi-trusted
config, model metadata, CLI overrides, or other externally controlled data to
hydra.utils.instantiate() without constraining which targets may be
instantiated.
Trusted application-owned configuration is not affected in the same way.
Remediation
Hydra 1.3.4 hardens the existing behavior by adding a blacklist of obvious dangerous targets. It is a substantial security improvement, and users remaining on the 1.3 release line should upgrade to 1.3.4 or a newer version.
The unreleased Hydra 1.4 development line uses an allowlist-based instantiation model that fully addresses this vulnerability class. The allowlist must come from trusted application code or another trusted channel, not from the untrusted config being instantiated.
Applications that consume untrusted or semi-trusted config should not pass it
directly to hydra.utils.instantiate(). They should validate _target_ values
against a trusted allowlist before instantiation.
{
"affected": [
{
"database_specific": {
"last_known_affected_version_range": "\u003c= 1.3.3"
},
"package": {
"ecosystem": "PyPI",
"name": "hydra-core"
},
"ranges": [
{
"events": [
{
"introduced": "0"
},
{
"fixed": "1.3.4"
}
],
"type": "ECOSYSTEM"
}
]
}
],
"aliases": [
"CVE-2026-68508"
],
"database_specific": {
"cwe_ids": [
"CWE-470",
"CWE-94"
],
"github_reviewed": true,
"github_reviewed_at": "2026-08-21T20:57:31Z",
"nvd_published_at": null,
"severity": "HIGH"
},
"details": "## Summary\n\n`hydra.utils.instantiate()` resolves and calls Python objects from config. If an\napplication passes untrusted config to `instantiate()`, an attacker who controls\n`_target_` and its arguments can cause arbitrary code execution in the consuming\nprocess.\n\nHydra is not a network service. Exploitation requires a consuming application,\nlibrary, or user workflow to load attacker-controlled config, CLI overrides, or\nmodel metadata and pass it to `hydra.utils.instantiate()`.\n\n## Details\n\nHydra\u0027s instantiate API is designed to construct objects and call functions from\nconfiguration. For example:\n\n```yaml\ncomponent:\n _target_: package.module.Class\n arg: value\n```\n\nWhen this config is passed to `hydra.utils.instantiate()`, Hydra resolves\n`_target_` and calls it with the provided arguments.\n\nThis is intended for trusted application configuration. However, if untrusted\ninput controls `_target_`, the config becomes a callable-selection mechanism. A\nmalicious config can select a callable capable of executing code or commands and\nprovide attacker-controlled arguments.\n\nThis issue is the same general class of problem discussed by Unit 42 for\ndownstream AI/ML libraries such as NVIDIA NeMo, where untrusted model metadata\nwas passed into Hydra instantiate:\n\nhttps://unit42.paloaltonetworks.com/rce-vulnerabilities-in-ai-python-libraries/\n\nHydra 1.3.4 includes a blacklist for some dangerous `_target_` values. That\nblacklist is defense-in-depth and is not a complete security boundary. The\nblacklist is not present in the released `hydra-core` 1.3.3 package, so this\nissue should not be described as a bypass of a released 1.3.3 blacklist.\n\n## Impact\n\nA successful attack can execute code in the process that calls\n`hydra.utils.instantiate()`. The impact is limited to the privileges and\nenvironment of that process.\n\nPotential impact includes:\n\n- Reading files, credentials, environment variables, or data accessible to the\n process\n- Modifying files, outputs, checkpoints, or application state writable by the\n process\n- Terminating or disrupting the process\n\n## Affected Usage\n\nApplications and libraries are affected when they pass untrusted or semi-trusted\nconfig, model metadata, CLI overrides, or other externally controlled data to\n`hydra.utils.instantiate()` without constraining which targets may be\ninstantiated.\n\nTrusted application-owned configuration is not affected in the same way.\n\n## Remediation\n\nHydra 1.3.4 hardens the existing behavior by adding a blacklist of obvious\ndangerous targets. It is a substantial security improvement, and users remaining\non the 1.3 release line should upgrade to 1.3.4 or a newer version.\n\nThe unreleased Hydra 1.4 development line uses an allowlist-based instantiation\nmodel that fully addresses this vulnerability class. The allowlist must come\nfrom trusted application code or another trusted channel, not from the untrusted\nconfig being instantiated.\n\nApplications that consume untrusted or semi-trusted config should not pass it\ndirectly to `hydra.utils.instantiate()`. They should validate `_target_` values\nagainst a trusted allowlist before instantiation.",
"id": "GHSA-2cp2-2r3c-7p7r",
"modified": "2026-08-21T20:57:31Z",
"published": "2026-08-21T20:57:31Z",
"references": [
{
"type": "WEB",
"url": "https://github.com/hydra-ecosystem/hydra/security/advisories/GHSA-2cp2-2r3c-7p7r"
},
{
"type": "WEB",
"url": "https://github.com/hydra-ecosystem/hydra/issues/3259"
},
{
"type": "WEB",
"url": "https://github.com/hydra-ecosystem/hydra/pull/3261"
},
{
"type": "WEB",
"url": "https://github.com/hydra-ecosystem/hydra/commit/7faad0dcedfb4c0a364aa1067c0080fd6fdf8dca"
},
{
"type": "PACKAGE",
"url": "https://github.com/hydra-ecosystem/hydra"
},
{
"type": "WEB",
"url": "https://github.com/hydra-ecosystem/hydra/releases/tag/v1.3.4"
}
],
"schema_version": "1.4.0",
"severity": [
{
"score": "CVSS:3.1/AV:L/AC:L/PR:N/UI:R/S:U/C:H/I:H/A:H",
"type": "CVSS_V3"
}
],
"summary": "Hydra: hydra.utils.instantiate with untrusted config can lead to code execution"
}
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.