The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill implements a custom _VocabUnpickler in scripts/_utils.py that restricts pickle.load to a specific allow-list of classes (TorchVocab, MolVocab, SMILESVocab, Counter, re.compile). This effectively mitigates the risks associated with unsafe deserialization of vocabulary files.
[DYNAMIC_EXECUTION]: The load_checkpoint function in scripts/_utils.py utilizes torch.load with weights_only=True and a strictly defined set of safe_globals (including NumPy multarray reconstructors and scalars). This follows security best practices for loading machine learning models to prevent arbitrary code execution.
[COMMAND_EXECUTION]: Several scripts (run_inference.py, prepare_data.py) use subprocess.run to orchestrate the internal workflow. These calls are used to invoke the core main.py entry point or utility scripts within the same repository. The command arguments are constructed from deterministic paths and validated user inputs, representing standard orchestration behavior rather than a security risk.
[PRIVILEGE_ESCALATION]: The Docker container management in scripts/kermt_container.sh explicitly uses the --user "$(id -u):$(id -g)" flag, ensuring that processes inside the container do not run with root privileges on the host system.
[DATA_EXFILTRATION]: Environment variable forwarding is limited to specific runtime configurations (e.g., CUDA settings, W&B API keys, and Hugging Face tokens). The runner_environment utility in scripts/_utils.py is designed to keep unrelated host credentials isolated from the containerized jobs.

kermt-infer