goal-loop
Pass
Audited by Gen Agent Trust Hub on Aug 13, 2026
Risk Level: SAFEPROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [PROMPT_INJECTION]: The orchestration framework ingests and processes data from external sources, which creates a potential surface for indirect prompt injection. 1. Ingestion points: User-defined goals (Phase 1), worker-provided evidence (Phase 3), and external verification signals such as logs or screenshots (Phase 4). 2. Boundary markers: The skill instructions lack technical delimiters (e.g., XML tags or specific delimiters) to isolate untrusted data, relying instead on role-based separation between workers and evaluators. 3. Capability inventory: The framework supports file system writes (durable ledger mode) and execution of shell commands through workers. 4. Sanitization: No specific routines for escaping or filtering content from external verification outputs are described.
- [COMMAND_EXECUTION]: The skill instructs the agent to execute tools and scripts to verify completion criteria during Phase 4 (External Verification). While the skill itself does not contain malicious code, the workflow is designed to invoke and capture the output of shell commands and capture exit codes to drive the goal-loop logic.
Audit Metadata