long-context
Warn
Audited by Gen Agent Trust Hub on Oct 29, 2026
Risk Level: MEDIUMEXTERNAL_DOWNLOADSREMOTE_CODE_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill suggests downloading code from individual GitHub repositories, including
https://github.com/jquesnelle/yarnandhttps://github.com/ofirpress/attention_with_linear_biases, which are personal accounts rather than trusted organizations. - [REMOTE_CODE_EXECUTION]: Instructions in
SKILL.mddirect the user or agent to clone a repository and immediately run a Python script (python scripts/train.py). This promotes the execution of unverified external code from a third-party source. - [INDIRECT_PROMPT_INJECTION]: The skill processes long documents (e.g., from datasets like
pg19) that may contain malicious instructions designed to influence the model's behavior. - Ingestion points: External datasets and user documents are consumed during training and inference in
SKILL.mdandreferences/fine_tuning.md. - Boundary markers: The implementation code lacks delimiters or explicit instructions to distinguish untrusted data from system prompts.
- Capability inventory: The skill enables model fine-tuning and text generation based on long-context inputs.
- Sanitization: No methods for sanitizing or filtering input text are implemented in the provided code snippets.
Audit Metadata