architect

Warn

Audited by Gen Agent Trust Hub on Oct 22, 2026

Risk Level: MEDIUMCOMMAND_EXECUTIONDYNAMIC_EXECUTIONPRIVILEGE_ESCALATIONINDIRECT_PROMPT_INJECTIONOBFUSCATION
Full Analysis
  • [COMMAND_EXECUTION]: The check-runner.ps1 and check-runner.sh scripts are designed to parse markdown documentation for commands (prefixed with - RUN: ) and execute them directly using Invoke-Expression (PowerShell) or bash -c (Bash).
  • [DYNAMIC_EXECUTION]: The check-runner.ps1 script dynamically generates a PowerShell wrapper for each check command at runtime, which is then executed as a sub-process.
  • [PRIVILEGE_ESCALATION]: The dispatch.md file contains explicit instructions for the agent to use the --sandbox danger-full-access flag when launching builder and strategist jobs. This intentionally opts out of default security constraints to bypass technical issues with restricted environments.
  • [INDIRECT_PROMPT_INJECTION]: The skill creates a vulnerability surface for indirect prompt injection through its core orchestration loop. 1. Ingestion points: The skill ingests untrusted data from GitHub issue comments and local job reports in docs/jobs/. 2. Boundary markers: While some dispatch templates use separators, the raw ingestion from trackers lacks robust boundary enforcement. 3. Capability inventory: The skill possesses high-impact capabilities including arbitrary command execution (check-runner), git operations (postflight), and process spawning (run-job). 4. Sanitization: There is no evidence of explicit input sanitization or validation of external content before it is processed by the agent.
  • [OBFUSCATION]: The check-runner.ps1 script utilizes Base64 encoding to wrap and pass commands to PowerShell's -EncodedCommand parameter, which can obscure the intent of the executed code from static analysis tools.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Oct 22, 2026, 09:35 PM