Model files are untrusted input.

• Parsers must bounds-check every offset and length
• Tensor dimensions and allocation sizes are overflow-checked
• PyTorch pickle payloads are not deserialized by default
• Embedded code in artifacts is never executed
• Do not store secrets in nn config files

Report vulnerabilities privately to the maintainers. Do not open a public issue with exploit details.