kermt-pretrain-scratch
Pass
Audited by Gen Agent Trust Hub on Aug 13, 2026
Risk Level: SAFEPROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [PROMPT_INJECTION]: The skill defines a workflow where shell commands are constructed using user-provided inputs, such as file paths and training hyperparameters.\n
- Ingestion points: User-supplied corpus CSV paths and command-line arguments (e.g.,
--epochs,--batch-size,--pretrain-target-mode) are interpolated into shell commands.\n - Boundary markers: The instructions do not specify any delimiters or safety warnings to prevent the agent from executing embedded instructions or special characters found in the user-provided strings.\n
- Capability inventory: The workflow involves executing
dockercommands via a local wrapper script (kermt_container.sh), which provides a vector for command injection if inputs are not sanitized.\n - Sanitization: No explicit sanitization or validation logic is defined for the user-supplied parameters before they are used in shell command templates.\n- [COMMAND_EXECUTION]: The skill performs multiple shell operations, including system probing (e.g.,
torch.cudachecks) and containerized execution of Python scripts. These operations are essential to the skill's primary function of training models but rely on the integrity of the local repository and the agent's ability to handle inputs safely.
Audit Metadata