| --- |
| library_name: transformers |
| tags: |
| - ctf |
| - model-scanning |
| - supply-chain-canary |
| - static-analysis |
| license: apache-2.0 |
| --- |
| |
| # WWTCyberLab/static-canary-pt-loader-risk |
|
|
| This is an inert CTF canary model repository for Cisco SecureAI model |
| scanning demos. It is intentionally designed to look risky to static |
| model scanners while avoiding destructive behavior. |
|
|
| Expected scanner themes: |
|
|
| - `model_CTF{model_scanner_high_pickle_evidence}.pt` contains a Python protocol-2 pickle reducer |
| for `__builtin__.set`. Loading it constructs only a normal set of |
| static strings, including `CTF_MODEL_CANARY`. |
| - `static-canary-pt-loader-risk*.safetensors` files are valid |
| safetensors fixtures with static metadata, suspicious tensor names, |
| anomalous FP32 markers, and deterministic high-entropy-looking |
| weight bytes to exercise obfuscation heuristics. |
| - `config.json` includes `auto_map` metadata so scanners can exercise |
| `trust_remote_code` review paths. |
| - `modeling_static_canary.py` imports modules commonly flagged by |
| loader-risk scanners, but import and model construction have no |
| subprocess, filesystem, or network side effects. |
|
|
| Runtime canary events should be emitted by the lab harness, not hidden |
| inside the model artifact. |
|
|
