advisor-orchestrator-worker
Warn
Audited by Gen Agent Trust Hub on Nov 5, 2026
Risk Level: MEDIUMINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADSDYNAMIC_EXECUTION
Full Analysis
- [COMMAND_EXECUTION]: The skill dispatches workers via the
agyCLI using shell command substitution$(cat "$brief")to pass prompt content. This pattern creates a potential shell injection surface if the content of the brief file contains malicious shell characters or control sequences, despite the skill's documentation explicitly warning against direct string interpolation. - [COMMAND_EXECUTION]: The worker dispatch logic utilizes the
--dangerously-skip-permissionsflag when invoking theagyCLI tool. This flag is designed to bypass the tool's internal security and permission model to facilitate execution in non-interactive environments, effectively removing a layer of access control. - [DYNAMIC_EXECUTION]: The orchestrator dynamically assembles and executes shell commands in background subshells to manage parallel worker tasks. This involves creating temporary directories and manipulating environment variables (
env -i HOME="$HOME" PATH="$PATH") to isolate execution contexts, though the core command assembly remains vulnerable to content-based injection. - [INDIRECT_PROMPT_INJECTION]: The skill serves as an orchestrator that processes untrusted data generated by external "worker" models and "advisor" models, creating an indirect prompt injection attack surface.
- Ingestion points: Reads task results from worker output files and strategy critiques from advisor consult files.
- Boundary markers: Implements execution isolation using
mktempdirectories, separate subshells, andjq --rawfileto safely load external content into JSON payloads for API calls. - Capability inventory: Executes arbitrary shell commands via multiple CLI tools (
agy,claude,jq,perl), performs network requests, and manages local file system operations. - Sanitization: Uses
jqfor structured data construction in the API fallback path, which prevents JSON injection, but relies on standard file redirection for the CLI execution paths. - [EXTERNAL_DOWNLOADS]: The skill performs network operations via
curlto reach well-known service endpoints including Google's Generative Language API (generativelanguage.googleapis.com) and Anthropic's API (api.anthropic.com) during its fallback routines. These connections are used to maintain functionality when local CLI tools are unavailable.
Audit Metadata