design-to-code

Pass

Audited by Gen Agent Trust Hub on Mar 23, 2026

Risk Level: SAFEPROMPT_INJECTIONCOMMAND_EXECUTIONDATA_EXFILTRATION
Full Analysis
  • [PROMPT_INJECTION]: Indirect Prompt Injection Surface. The skill ingests untrusted design specifications or references through the $ARGUMENTS variable, which then steers the autonomous modification of the codebase and execution of build tools across four sequential phases.\n
  • Ingestion points: The $ARGUMENTS variable defined in SKILL.md serves as the entry point for external data.\n
  • Boundary markers: There are no explicit delimiters or instructions to the agent to ignore embedded commands within the input arguments.\n
  • Capability inventory: The skill has the capability to write files, commit code, and execute shell commands (build, test) across multiple development phases.\n
  • Sanitization: No sanitization or validation of the input arguments is performed before the data is used to drive agent actions.\n- [COMMAND_EXECUTION]: Automated execution of build and test tools. The 'Self-Healing Validation' section directs the agent to execute shell commands to compile code and run tests based on the generated output.\n
  • Evidence: The skill explicitly instructs the agent to 'Run the project's test suite' and 'Run build/compile' up to three times to validate fixes.\n- [DATA_EXFILTRATION]: Local interaction with internal application metadata. The 'Self-Evolution Telemetry' phase involves reading from and writing to a hidden platform-specific directory in the user's home folder.\n
  • Evidence: The skill logic searches for the project path in ~/.claude/projects/ and appends execution logs to skill-telemetry.md. While this is intended for self-improvement, it involves autonomous access to internal configuration directories.
Audit Metadata
Risk Level
SAFE
Analyzed
Mar 23, 2026, 10:58 AM
Security Audit — agent-trust-hub — design-to-code