sync-codex
Pass
Audited by Gen Agent Trust Hub on Nov 4, 2026
Risk Level: SAFECOMMAND_EXECUTIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTIONPROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The orchestrator script (
run-codex-sync.mjs) manages the sequential execution of approximately 19 local scripts and tests usingchild_process.spawn. This creates a centralized entry point for executing arbitrary code within the repository bundle. - [DYNAMIC_EXECUTION]: The script uses dynamic module loading via
requirewith paths computed at runtime based on the repository root (e.g., loadingworkflow-skill-contract.cjs). It also dynamically constructs command-line arguments and environment variables for its sub-processes. - [INDIRECT_PROMPT_INJECTION]: The skill functions as a data pipeline that ingests repository files like
CLAUDE.mdto generate downstream agent instructions inAGENTS.mdand skill mirrors. This creates a vector where malicious instructions in a repository source file could influence the behavior of agents interacting with the project. - Ingestion points:
CLAUDE.md,.claude/skills/**, and project configuration files (e.g.,.ck.json). - Boundary markers: The pipeline does not implement explicit boundary markers or "ignore embedded instructions" warnings when mirroring natural language content into generated instruction files.
- Capability inventory: The orchestrator has the capability to execute repository scripts (
spawn), read and write project files, and hand off execution to secondary sync pipelines. - Sanitization: There is no evidence of sanitization or safety validation for the instructions being synced between source and generated files.
- [PROMPT_INJECTION]: The
SKILL.mdinstructions include a 'Subagent authorization' clause designed to override standard authorization protocols. It instructs the agent that invoking the skill automatically 'authorizes' the use ofspawn_agenttools for associated tasks, which attempts to bypass manual authorization checkpoints.
Audit Metadata