experiment-report-writing
Pass
Audited by Gen Agent Trust Hub on Oct 30, 2026
Risk Level: SAFEDATA_EXFILTRATIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [DATA_EXFILTRATION]: The skill systematically collects sensitive system and project-level information, including exact commands, execution logs (stdout/stderr), environment configurations, machine specifications (GPU, CUDA, drivers), and code provenance (branches, commits). This data is then formatted and sent to external platforms such as Notion, Feishu, or Obsidian. While these are well-known services, the inclusion of generic configuration and log data poses a risk of exposing secrets or environment variables stored in those files.
- [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and summarize content from potentially untrusted external files, which could contain malicious instructions designed to influence the agent's behavior during the report generation process.
- Ingestion points: The agent is instructed to inspect stdout/stderr logs, TensorBoard/W&B tables, evaluation JSON/CSV files, and various configuration scripts as listed in the 'Evidence Intake' section of SKILL.md.
- Boundary markers: The skill uses Markdown structural elements (e.g., code blocks for metadata) but does not provide explicit instructions for the agent to distinguish between data and instructions within the ingested logs.
- Capability inventory: The collected data is processed and written to external documentation tools via tool-chaining (e.g., notion-doc-workflow).
- Sanitization: There is no evidence of sanitization, filtering, or escaping for the contents of the logs or artifacts before they are interpolated into the final report.
Audit Metadata