lora-training
Pass
Audited by Gen Agent Trust Hub on Nov 4, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes external data such as image filenames and text captions provided in a user-specified folder. This constitutes an indirect prompt injection surface. However, the risk is mitigated by a mandatory human-in-the-loop approval gate. The skill uses
scripts/dataset_check.pyto hash the entire dataset andscripts/aitoolkit_config.pyverifies this hash before generating configurations, ensuring no changes were made to the data after the user approved it. - Ingestion points:
scripts/dataset_check.py(viaos.listdirandopen()) and theffmpegcommand instruction inSKILL.md. - Boundary markers: The skill explicitly instructs the agent to wait for human approval of a contact sheet and warnings before proceeding. The
APPROVED.jsonfile serves as a cryptographic boundary marker for the dataset state. - Capability inventory: File system read/write (JSON and YAML), and shell command execution (
ffmpeg) as instructed in the documentation. - Sanitization: The
aitoolkit_config.pyscript usesjson.dumps()when generating YAML, which provides safe escaping of potentially malicious characters in user-provided names or trigger words. - [COMMAND_EXECUTION]: The skill instructions provide shell command templates, such as
ffmpeg, for generating contact sheets. These templates incorporate user-supplied folder paths. While this is a common development pattern, it relies on the agent's ability to properly escape shell metacharacters in the folder path to prevent command injection. - [SAFE]: The Python scripts
scripts/dataset_check.pyandscripts/aitoolkit_config.pyare self-contained utilities that use only Python's standard library. They perform no network requests, use no external dependencies, and do not execute shell commands directly viasubprocessoros.system. They are focused on data validation, hashing, and configuration templating.
Audit Metadata