Vulnerability  ·  2026-09-22

Feast ML feature store: pull_request_target workflow leak of GCP/AWS/Snowflake credentials (CVSS 8.9)

VulnerabilityHigh impactGlobalCVE-2026-55563
Published 2026-09-21. The stale-label + pull_request_target pattern lets an untrusted fork contributor run attacker code in a privileged CI context holding ML-infrastructure cloud credentials. CISA SSVC assessed exploitation as 'poc' with total technical impact.
ML-infrastructure CI is where model training/Serving credentials live; a fork contributor gaining them via the merge pull_request_target gives supply-chain style access to model training data, feature stores and downstream cloud resources of the Feast project's deployment.
.github/workflows/pr_integration_tests.yml uses pull_request_target with the synchronize event and preserves ok-to-test/approved/lgtm labels across newly pushed commits. A fork contributor obtains approval on a benign revision, then pushes changed code that is executed through privileged make targets from the merge ref, exposing cloud credentials and enabling runner code execution and downstream cloud access.
feast-dev/feast < 0.65.0
Upgrade to Feast v0.65.0 (commit 76192229; GHSA-2j2x-r73g-hrr5); projects using similar workflows should remove stale approval labels on commit push.
NVD CVE-2026-55563Feast advisory GHSA-2j2x-r73g-hrr5
See this in the live feed Explore related AI security and governance findings — updated every morning.
Open the feed →