novel-research

Pass

Audited by Gen Agent Trust Hub on Sep 24, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONOBFUSCATIONCOMMAND_EXECUTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill implements a feature to import character metadata from external PNG and WebP images (SillyTavern character cards). This represents an indirect prompt injection surface where malicious instructions could be embedded in the image metadata to influence the agent's behavior.
  • Ingestion points: The scripts/parse-charcard.mts script reads EXIF/tEXt/iTXt data from user-provided image files in the Phase 3 interview stage.
  • Boundary markers: The SKILL.md file (Phase 3) and references/style-analysis-guide.md contain explicit warnings to the agent: "不可直接注入小说写作上下文" (Do not directly inject into the novel writing context) and "分析其意图 → 过滤角色扮演框架" (Analyze intent → filter roleplay frameworks).
  • Capability inventory: The skill uses the node environment to execute local TypeScript-based scripts (parse-charcard.mts, validate-research-project.mts) and has the capability to write various Markdown files and create directories within the user's workspace.
  • Sanitization: The scripts/lib/charcard-transformer.mts utility formats imported data into structured Markdown sections and wraps sensitive instructions (like raw system prompts from the imported card) in warning blocks and comments to ensure the agent treats them as data rather than instructions.
  • [OBFUSCATION]: The skill includes a minified third-party vendor library, scripts/lib/vendor/exifreader.js. While minification is a standard practice for reducing file size in web development, it obscures the code's logic from direct review. Static analysis tools flagged this file for containing Base64 encoding and decoding logic, which is used here for its intended purpose of processing binary image metadata.
  • [COMMAND_EXECUTION]: The skill relies on executing local scripts using node --experimental-strip-types. These scripts perform file system operations, such as reading templates, creating project directories, and writing project files. The execution is strictly scoped to the skill's own directory and the designated project root provided by the user.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 24, 2026, 06:37 AM