create-table-classifier

SUSPICIOUS: The skill's purpose and visible capabilities mostly align with local ML training, and the only named external service is official W&B telemetry. The main concern is execution trust: all meaningful behavior is hidden behind an unprovided local Bash script with no provenance, install transparency, or integrity information, so the real data flows cannot be fully verified.

Code is functionally benign for predicting table extraction strategies and exporting model artifacts, but it exposes a significant supply-chain risk: torch.load is used to unpickle model checkpoints without validation or integrity checks. Loading model files from untrusted sources could result in arbitrary code execution. Recommend: treat .pth files as untrusted, implement signature/hmac verification or use torch.load with restricted/untrusted handlers if available, avoid writing absolute exported_from paths or sanitize them, and consider running model loading in an isolated environment. No direct evidence of malware or obfuscation in this file itself.

The script contains no direct malicious payloads, but it exposes significant supply-chain and execution-time risks: automatic pip installs and loading remote models with trust_remote_code=True permit arbitrary remote code execution in the runtime. Combined with writing model artifacts to disk, this creates potential for compromise if untrusted model repos or packages are used. Treat use in untrusted environments as risky and follow mitigations (no auto-installs, avoid trust_remote_code, use pinned artifacts, run in isolated environments).