plankton-code-quality

Warn

Audited by Gen Agent Trust Hub on Mar 6, 2026

Risk Level: MEDIUMEXTERNAL_DOWNLOADSREMOTE_CODE_EXECUTIONCOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
  • [EXTERNAL_DOWNLOADS]: The installation process requires cloning an external, untrusted repository (github.com/alexfazio/plankton.git) and integrating its shell scripts as automated hooks. These scripts are executed with the user's privileges and originate from a source outside the trusted list.
  • [REMOTE_CODE_EXECUTION]: The skill implements a 'Phase 3: Delegate + Verify' mechanism that spawns claude -p subprocesses. These subprocesses receive JSON-formatted linter violations and are instructed to automatically modify source code, which constitutes automated code execution and modification based on potentially untrusted input.
  • [COMMAND_EXECUTION]: The skill utilizes multiple shell hooks (multi_linter.sh, protect_linter_configs.sh, stop_config_guardian.sh) that run automatically on every file edit or command attempt. These hooks perform a wide range of filesystem and system operations.
  • [PROMPT_INJECTION]: The skill presents an indirect prompt injection surface (Category 8) because it processes untrusted source code and feeds linter output into a secondary LLM process.
  • Ingestion points: Source code files modified by the user or agent, which are then read by various linters.
  • Boundary markers: None are specified; the violations JSON is passed directly to the claude -p subprocess.
  • Capability inventory: The subprocess has write access to the filesystem to perform code fixes.
  • Sanitization: There is no evidence that linter error messages (which can include content from the source code itself) are sanitized before being processed by the secondary model.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Mar 6, 2026, 02:30 AM