architect-research

Pass

Audited by Gen Agent Trust Hub on Jul 15, 2026

Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSPROMPT_INJECTIONDATA_EXFILTRATION
Full Analysis
  • [COMMAND_EXECUTION]: The skill orchestrates sub-agents by executing the codex exec command in the shell. It employs security best practices, such as using stdin for prompt delivery to avoid shell injection and applying a --sandbox read-only flag to the execution environment.
  • [EXTERNAL_DOWNLOADS]: Researchers are directed to query numerous well-known technology and academic services, including GitHub, ArXiv, Semantic Scholar, HuggingFace, and NPM, to gather technical evidence.
  • [PROMPT_INJECTION]: The skill is exposed to indirect prompt injection via the ingestion of untrusted web content. It implements extensive mitigation strategies, including a 'findings discipline' that prioritizes primary sources, a requirement for at least two independent sources per claim, and an adversarial verification pass.
  • [DATA_EXFILTRATION]: The orchestration logic reads local configuration files from ~/.architect/config and project-local directories to determine execution settings and tool defaults.
Audit Metadata
Risk Level
SAFE
Analyzed
Jul 15, 2026, 10:07 AM