add-model-eval-config
Pass
Audited by Gen Agent Trust Hub on Nov 5, 2026
Risk Level: SAFEREMOTE_CODE_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [REMOTE_CODE_EXECUTION]: The skill facilitates the installation of the NeMo Gym framework by specifying a remote Git repository (https://github.com/NVIDIA-NeMo/gym) and a specific commit reference in the driver.gym_install configuration block. This download originates from a verified vendor repository.
- [INDIRECT_PROMPT_INJECTION]: The skill instructions direct the agent to retrieve model parameters (context length, sampling settings, and parsers) by reading and parsing content from model checkpoint files, such as chat_template.jinja, config.json, and generation_config.json. Since these files may be sourced from external parties, they represent an attack surface for indirect prompt injection.
- Ingestion points: File content from chat_template.jinja, tokenizer_config.json, generation_config.json, and config.json located within the model checkpoint directory.
- Boundary markers: The skill does not provide specific instructions or delimiters to isolate the data read from these files or to ignore potential instructions embedded within them.
- Capability inventory: The agent has the capability to generate and submit job configurations (gym eval submit) that execute shell commands and install software from remote repositories.
- Sanitization: There are no explicit steps for sanitizing or validating the data retrieved from the checkpoint files before it is incorporated into the generated serving configurations.
Audit Metadata