Executive brief
Weights & Biases (wandb) is a platform used by AI developers to track machine learning experiments and manage data artifacts. A security flaw in the way the software validates data integrity could allow an attacker to bypass checks that ensure files haven't been tampered with. This could potentially lead to the use of modified or malicious data during model training or deployment, though the attack is difficult to execute and requires specific conditions.
Technical details
A vulnerability exists in the Artifact Integrity Validation component of wandb (specifically within ArtifactManifestEntry.download in wandb/sdk/lib/hashutil.py). The system relies on a weak or non-collision-resistant hash (MD5) for verifying the integrity of downloaded artifacts. A remote attacker with low privileges could potentially exploit this by substituting legitimate artifact content with malicious data that produces a hash collision. While the attack complexity is high and exploitability is considered difficult, a proposed fix involves migrating to SHA-256 for content digests to ensure robust integrity checks.
Affected products
- Weights & Biases wandb 0.25.2.dev1
Timeline
- 2026-06-11: disclosed: Issue reported and PR created on GitHub
- 2026-07-13: advisory: CVE published and NVD record created