quickscan

Warn

Audited by Gen Agent Trust Hub on May 10, 2026

Risk Level: MEDIUMCOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill uses uv run python3 to execute multiple local scripts within the project directory. It passes $ARGUMENTS directly to the shell, which can be exploited for arbitrary command execution.
  • Evidence: uv run python3 $CLAUDE_PROJECT_DIR/tools/brain.py brief $ARGUMENTS in SKILL.md.
  • Evidence: uv run python3 $CLAUDE_PROJECT_DIR/tools/scope_check.py $ARGUMENTS in SKILL.md.
  • [PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection because it ingests untrusted data into sensitive command-line operations.
  • Ingestion points: User-provided input via $ARGUMENTS in SKILL.md.
  • Boundary markers: Absent. The input is not wrapped in quotes or delimiters to prevent shell metacharacter expansion.
  • Capability inventory: Subprocess execution via uv run, interaction with brain.py (database/logging), and the ability to launch sub-agents (recon, config-auditor).
  • Sanitization: Absent. The instructions do not define any validation or escaping for the external content before it is used in a shell context.
Audit Metadata
Risk Level
MEDIUM
Analyzed
May 10, 2026, 07:58 AM
Security Audit — agent-trust-hub — quickscan