ZDI Publishes 47 Meta AI Zero-Days

This is a major supply-chain warning for ML developers: model artifacts and configuration files are executable inputs when loaders trust unsafe serialization formats.

• –The disclosures span 47 advisories, all rated CVSS 7.8, across three major Meta-maintained AI libraries.
• –Exploitation generally requires opening a malicious file or visiting a malicious page, making shared research and training environments especially exposed.
• –Developers should treat checkpoints, configs, and datasets as untrusted; sandbox loaders and eliminate unsafe pickle/YAML deserialization paths.
• –The incident highlights the security cost of fragmented maintenance across popular open-source research frameworks.