unified-execute-with-file

Warn

Audited by Gen Agent Trust Hub on Mar 25, 2026

Risk Level: MEDIUMCOMMAND_EXECUTIONREMOTE_CODE_EXECUTIONPROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill executes arbitrary bash commands retrieved from the 'convergence.verification' field of external JSON task files, enabling arbitrary command execution on the host system.
  • Evidence: SKILL.md Phase 3, Step 3.2 uses 'Bash(verification)' to run the command specified in the JSON.
  • [REMOTE_CODE_EXECUTION]: The execution logic permits network-enabled tools like 'curl' to be executed if present in the task data, facilitating potential remote script execution.
  • Evidence: The 'isExecutableCommand' helper function explicitly includes 'curl' in its whitelist.
  • [PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it ingests untrusted task data and uses it to guide the agent's editing and writing behavior without sanitization or boundary markers.
  • Ingestion points: .task/*.json files processed in Phase 1.
  • Boundary markers: Absent; task descriptions and criteria are interpolated directly into the execution loop.
  • Capability inventory: Bash, Read, Write, Edit, Grep, Glob, AskUserQuestion.
  • Sanitization: Absent for descriptive fields and minimal regex-based validation for verification commands.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Mar 25, 2026, 11:06 AM
Security Audit — agent-trust-hub — unified-execute-with-file