pyvene-interventions
Pass
Audited by Gen Agent Trust Hub on Oct 5, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSDATA_EXFILTRATIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill references the installation of the 'pyvene' package from a standard public registry. It also demonstrates loading pre-trained models and intervention configurations from the Hugging Face Hub, including specific community repositories.\n
- Evidence: 'pip install pyvene' and 'pv.IntervenableModel.load("zhengxuanzenwu/intervenable_honest_llama2_chat_7B", ...)' in SKILL.md and references/README.md.\n- [DATA_EXFILTRATION]: The skill documentation includes examples of sharing research results by uploading intervention configurations to an external hub for collaboration.\n
- Evidence: 'intervenable.save_intervention("username/my-intervention")' in SKILL.md and references/api.md.\n- [INDIRECT_PROMPT_INJECTION]: The skill defines workflows that involve processing user-supplied prompts to observe model behavior, which constitutes an injection surface. This is an inherent part of the model analysis tasks described.\n
- Ingestion points: Prompts such as 'clean_prompt' and 'base_prompt' in SKILL.md and references/tutorials.md.\n
- Boundary markers: None present in the provided code snippets.\n
- Capability inventory: Capabilities include writing files to disk via 'save()' and network communication for external repository interaction.\n
- Sanitization: The provided examples do not include explicit input sanitization, which is typical for research code.
Audit Metadata