Serialized model files such as pickles can execute code when loaded, and fine-tuned weights can carry hidden backdoors that survive safety training. Prefer safe formats, scan artifacts before loading them, and verify where they came from.
Read in this order
- OfficialHugging Face: Pickle ScanningStart here: why pickles are dangerous and how model hubs scan for them.
- Open sourcetrailofbits/fickling: A Python pickling decompiler and static analyzerA tool to inspect and analyze pickle files before you load them.
- PaperSleeper Agents: Training Deceptive LLMs that Persist Through Safety TrainingShows backdoored behavior can persist through safety training, so provenance matters.