← sources

How would you check if a given model weights is "safe" ?

source: reddit (reddit:LocalLLaMA)published 2026-10-25 14:16ingested 2026-10-26 02:40

retained text

Imagine you want to download a new model but it's recent and there is not consensus about wether its safe or not.

How would you check against potential malicious weights ?

There is the first "obvious" layer where the weights try to execute arbitrary code, but safetensors are for this. I am rather talking about less obvious things like template backdoor, llm poisoning etc

document fetched 2026-10-26 00:38 · canonical https://reddit.com/r/LocalLLaMA/comments/1xf3i59/how_would_you_check_if_a_given_model_weights_is