short-drama-novel-analyze
Pass
Audited by Gen Agent Trust Hub on Oct 7, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted novel content, creating a potential surface for indirect prompt injection.
- Ingestion points: The skill reads external novel files (e.g., "舓艥/舂舑舕艠舀舂興舃舑舕舂.txt") as described in "SKILL.md" and "novel_index.py".
- Boundary markers: Robust boundary instructions are present in "references/stage-contract.md" and "SKILL.md", prohibiting the copying of raw text and requiring the use of short, localized quotes with source spans.
- Capability inventory: Capabilities are limited to local Python script execution for indexing and writing markdown/JSONL analysis files to defined project directories. No network access or unsafe command evaluation is permitted.
- Sanitization: The skill enforces structured summaries and functional extraction rather than verbatim reproduction, minimizing the risk of instructions embedded in the novel influencing downstream agents.- [COMMAND_EXECUTION]: The skill executes local Python scripts to perform heavy data processing tasks.
- Evidence: "SKILL.md" invokes "scripts/novel_index.py" for indexing ("index"), verification ("verify"), and sampling ("sample"). The Python script itself uses standard libraries and performs safe, atomic file writes using "os.replace" and "os.fsync".- [SAFE]: No evidence of credential theft, data exfiltration, obfuscation, or persistence mechanisms was found. The skill maintains a clear separation between analysis (transformative) and decision-making (creative), reducing the risk of unauthorized autonomous actions.
Audit Metadata