LayerFault Security Corpus

The corpus contains deliberately synthetic, malformed and adversarial model repositories designed to reproduce security conditions that defensive tooling should detect.

Coverage includes:

• unsafe Pickle and serialization
• NumPy object arrays
• Safetensors structural anomalies
• PyTorch and TorchScript packages
• Hugging Face custom loaders and auto_map
• tokenizer and template attacks
• hidden and Unicode instructions
• Python, PowerShell, JavaScript and shell capability chains
• executable and native library payloads
• archive and path traversal attacks
• suspicious sidecars
• dependency and package integrity anomalies
• malformed model formats
• provenance and lineage mismatches
• synthetic weight anomalies and backdoor fixtures
• dataset poisoning
• runtime security test cases

The corpus is organised into Hugging Face Collections so individual security domains can be explored and tested independently.