alchemist-playbook
Pass
Audited by Gen Agent Trust Hub on Sep 14, 2026
Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill uses a provided local utility script,
scripts/estimate.py, to calculate training metrics such as FLOPs, compute budgets, and batch size math. This script runs locally without external dependencies or network access. - [EXTERNAL_DOWNLOADS]: The instructions and reference files cite numerous official research papers and repositories from reputable organizations including the Allen Institute for AI (AI2), Shanghai AI Lab, Tsinghua University, and Anthropic. These references are used solely for evidence-based technical advice.
- [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and analyze user-supplied training data, logs, and curves to diagnose training issues. While this processes external data, it is the primary function of the diagnostic tool and includes no execution pathways for malicious instructions embedded in logs.
- [SAFE]: The experiment tracking system utilizes self-contained HTML templates that operate entirely within the user's local environment or private repositories, avoiding the use of external CDNs or remote scripts to ensure data privacy.
Audit Metadata