model-training
Pass
Audited by Gen Agent Trust Hub on Sep 17, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process external datasets from user-defined locations, which introduces a surface for indirect prompt injection attacks.
- Ingestion points: Step 1 of the workflow in
SKILL.mdinvolves loading and profiling data from disk, databases, or remote storage. - Boundary markers: The instructions do not include boundary markers or delimiters to separate untrusted data from the agent's instructions, nor do they instruct the agent to ignore instructions embedded within the data.
- Capability inventory: The skill demonstrates file-system write capabilities, such as saving model weights and checkpoints (e.g.,
torch.saveandtrainer.save_modelinSKILL.md). - Sanitization: The workflow lacks explicit steps for sanitizing or validating the content of ingested data to prevent embedded malicious prompts from influencing the agent's behavior during data inspection or training.
Audit Metadata