SafeWeights GitHub

Cybersecurity · Hugging Face intake · Python 3.14

SafeWeights

Scan AI models before they reach your servers — statically, clearly, without running them.

Get started See the pipeline
0 model code executed
0 weight formats covered
0 sec to grasp the report

Intake pipeline

One path from approval request to a handoff-ready report.

What gets checked

Hover a format to see how SafeWeights inspects it.

Pickle / PyTorch

.pkl · .pt · .pth · .bin

Opcode & import scan via pickletools — flags dangerous GLOBAL / STACK_GLOBAL patterns.

SafeTensors

.safetensors

Validates header length, JSON metadata, and tensor offset bounds without loading weights.

GGUF

.gguf

Checks GGUF magic, version, and implausible tensor / metadata counts.

ONNX

.onnx

Protobuf heuristics plus scans for suspicious embedded markers.

Keras H5

.h5 · .keras

HDF5 magic + attribute walk for Lambda / exec-style payloads.

Urgent
Serious
Worth a look
Info

Reports people actually read

Built for non-technical teammates first — specialists get details below the fold.

PASS — Looks safe to proceed

Risk level: No serious problems found.

  • In 30 seconds — what we checked, what we found
  • Do this next — clear handoff steps
  • What stands out — plain-language cards
  • Details for specialists — file tables when needed

Run it on the VE

Python 3.14 · CustomTkinter · huggingface_hub

cd SafeWeights
python -m pip install -r requirements.txt
python main.py

Login uses the Hugging Face device link (same idea as hf auth login). Then pull a model ID, scan, and export the Markdown report.