tech-writer

Warn

Audited by Gen Agent Trust Hub on Mar 12, 2026

Risk Level: MEDIUMDATA_EXFILTRATIONPROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
  • [DATA_EXFILTRATION]: The skill's project scanning configuration, located in references/document-project/documentation-requirements.csv, explicitly instructs the agent to locate and analyze sensitive files. Targeted patterns include .env*, *auth*, *session*, *secret*, *keychain*, google-services.json, *jwt*, and *oauth*. This behavior leads to the extraction and exposure of potentially sensitive configuration details and credentials within the generated markdown documentation files.
  • [PROMPT_INJECTION]: The skill provides an Update Standards workflow that allows users to modify the references/documentation-standards.md file by adding "User Specified CRITICAL Rules". These rules are designed to supersede the agent's general instructions, creating a vector for modifying agent behavior through persistent prompt changes.
  • [COMMAND_EXECUTION]: The documentation workflow requires extensive file system access across the project workspace. The deep-dive-instructions.md specifies that the agent must read "every line of every file in scope" to identify implementation details, side effects, and logic, which constitutes a deep scan of the user's codebase.
  • [PROMPT_INJECTION]: The skill presents a surface for indirect prompt injection as it ingests and processes untrusted data from the user's codebase (Step 11 in full-scan-instructions.md). It extracts content from code and comments and interpolates it into documentation templates without explicit sanitization.
  • Ingestion points: Full-file reads of source code and comments during "Deep" and "Exhaustive" scans in full-scan-instructions.md and deep-dive-instructions.md.
  • Boundary markers: None implemented to distinguish codebase content from agent instructions in the output documents.
  • Capability inventory: The skill uses file-read and file-write tools to scan the repository and produce documentation.
  • Sanitization: There are no instructions for validating, escaping, or filtering the extracted codebase content before it is processed by the LLM or written to disk.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Mar 12, 2026, 09:05 PM