Glossary term
Glossary term
Security
Compromise of model weights, tuning artifacts, adapters, or deployment packages so the AI system behaves incorrectly, maliciously, or unreliably. Model artifact integrity should be protected through signing, secure repositories, restricted deployment rights, dependency scanning, and environment segregation.
Protect AI's Huntr platform documented multiple compromised models on Hugging Face in 2023 and 2024 containing malicious pickle deserialization payloads.
JFrog Security Research disclosed dozens of malicious models on Hugging Face in February 2024 capable of executing code on download.
Sigstore-based model signing initiatives at the Linux Foundation address model integrity verification through cryptographic signing.