kermt-pretrain-scratch

Pass

Audited by Gen Agent Trust Hub on Sep 17, 2026

Risk Level: SAFECOMMAND_EXECUTIONPRIVILEGE_ESCALATIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill utilizes subprocess.run across several utility and orchestration scripts (scripts/_utils.py, scripts/prepare_data.py, scripts/run_pretrain_local.py) to interface with local development tools such as git and docker. This execution is a core component of the skill's functionality to manage model training pipelines.
  • [PRIVILEGE_ESCALATION]: Execution depends on docker, which often requires elevated system permissions (e.g., membership in the docker group). The skill mitigates common container security risks by mapping the container user to the host's UID/GID and mounting host volumes as read-only where possible.
  • [DYNAMIC_EXECUTION]: In scripts/_utils.py, the skill implements a restricted pickle.Unpickler (_VocabUnpickler) to load vocabulary files. This implementation limits unpickling to a small set of safe, predefined classes (collections.Counter, re.compile), preventing the execution of arbitrary code during data loading.
  • [INDIRECT_PROMPT_INJECTION]: The skill provides an attack surface by ingesting and processing untrusted CSV datasets provided by users. These files are parsed by RDKit for chemical validation. While this is an inherent risk in data-processing tools, the skill uses established libraries for validation and limits the scope of the ingested data to training context.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 17, 2026, 11:42 AM
Security Audit — agent-trust-hub — kermt-pretrain-scratch