Hydra patches a code execution bug in instantiate()

Hydra's instantiate() function runs arbitrary code named in a config's target field, a flaw GitHub scores 7.8 out of 10 on the CVSS scale. The advisory tracks it as CVE-2026-68508 and covers every release up to 1.3.3.

Any app that loads config from an untrusted source can get hijacked. Instead of building the object it asked for, it runs someone else's code. Hydra isn't a network service on its own. The advisory says you need something layered on top that feeds instantiate() a value an attacker controls. That could be a config file, a CLI override, or metadata bundled with a downloaded model.

That third path is the one worth checking. Agent pipelines routinely call instantiate() on components described in a config shipped alongside a downloaded checkpoint. The checkpoint's author doesn't have to be trustworthy for that to work. Nothing in the normal flow stops target from naming a function that does more than build a model.

The fix is a version bump to 1.3.4. GitHub's advisory doesn't say whether the patch validates target against an allowlist or just changes how instantiate() resolves imports. That's worth checking before assuming a partial config still works.