CVE-2026-106439: Hydra: Mutable instantiate policy sets allow target blocklist bypass

Summary

Hydra's legacy instantiate() target blocklists and related execution-policy collections are stored in mutable module-level state. Because locate() can resolve attributes on imported objects, a configuration can resolve a mutation method such as .discard(), modify the active policy, and then instantiate a target that would otherwise be blocked.

Impact

A configuration controlling multiple sibling target entries can first remove an entry from a target blocklist and then invoke the removed target. Sibling nodes are processed in insertion order and consult the same mutable module-level policy.

This affects the legacy/default path without an execution whitelist. The 1.3 blocklist is a defense-in-depth measure rather than a complete security boundary, and applications must not treat arbitrary untrusted configuration as safe to instantiate or use for Python logging configuration.

Released hydra-core versions 1.3.4 through 1.3.6 and 1.4.0.dev4 through 1.4.0.dev9 are affected. Fixed releases are 1.3.7 and 1.4.0.dev10. The reported direct mutation path does not bypass an execution whitelist restricted to intended application targets and supplied by trusted Python code. During remediation, Hydra additionally hardened generic discovery, dispatch, introspection, alias, callable-result, and deferred-callable paths that could otherwise undermine name-only authorization.

Technical details

In hydra-core 1.3.4 and 1.3.5, the mutable blocklist is reachable as:

hydra.internal.instantiate.instantiate2.DEFAULTBLOCKLISTEDMODULES

In hydra-core 1.3.6, the expanded policy includes mutable collections in:

hydra.internal.targetpolicy

For example:

hydra.internal.targetpolicy.UNCONTROLLEDEXECUTIONTARGETS.discard

resolves to the bound set.discard method. The mutation target itself is not blocked on the legacy path. Once an entry is removed, subsequent authorization checks observe the modified set.

Other runtime policy collections can be attacked similarly by removing denied entries or adding entries to exception sets. Because the collections are module-level state, a successful mutation persists for the lifetime of the Python process unless explicitly reversed.

Safe reproduction

The behavior in hydra-core 1.3.6 can be demonstrated without invoking a shell command. The finally block restores the modified process-global state:

python from omegaconf import OmegaConf

from hydra.internal.targetpolicy import UNCONTROLLEDEXECUTIONTARGETS from hydra.utils import instantiate

target = "builtins.eval" assert target in UNCONTROLLEDEXECUTIONTARGETS

try: result = instantiate( OmegaConf.create( { "disarm": { "target": ( "hydra.internal.targetpolicy." "UNCONTROLLEDEXECUTIONTARGETS.discard" ), "args": [target], }, "proof": { "target": target, "args": ["40 + 2"], }, } ) )

assert result["proof"] == 42 assert target not in UNCONTROLLEDEXECUTIONTARGETS finally: UNCONTROLLEDEXECUTIONTARGETS.add(target)

Remediation

The fix makes runtime policy state immutable and integrity checked, and prevents declarative configuration from accessing or mutating Hydra internals and protected Python implementation state. Target authorization now covers canonical resolved identities, aliases, discovery results, callable results, deferred callables, and runtime arguments.

The patch also rejects configuration-driven code, policy, and process- environment mutation, along with unsafe introspection and formatting traversal that can expose protected runtime capabilities.

Hydra 1.3.7 receives these protections as defense in depth; it does not make untrusted configuration sandboxed. Hydra 1.4 additionally uses a trusted, narrowly scoped execution whitelist as the supported security boundary for declarative instantiation and Hydra-controlled Python logging configuration.

Users should upgrade to hydra-core 1.3.7 on the stable line or 1.4.0.dev10 on the development line.

Workarounds

Do not pass configuration from untrusted sources to instantiate() or to Hydra-controlled Python logging configuration. The 1.3 release line has no execution-whitelist facility, so users who cannot upgrade immediately must restrict configuration input to trusted sources.

On affected 1.4 development releases, applications can also supply a trusted, narrowly scoped execution whitelist from Python code. The whitelist itself must not be derived from untrusted configuration.

Restart any long-running process that may already have instantiated untrusted configuration, because a policy mutation persists in process-global state.