codex-orchestrator
Pass
Audited by Gen Agent Trust Hub on Oct 26, 2026
Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill relies on shell and Python scripts to interface with the
codexCLI and other system utilities (e.g.,script,jq,grep). These commands are used to execute agent tasks, process output, and manage autonomous goal runs. - [EXTERNAL_DOWNLOADS]: The
codex-version-check.shscript facilitates updates to the official@openai/codexpackage from the NPM registry. This is a standard update mechanism for the CLI tool the skill orchestrates and targets a well-known, trusted service. - [DYNAMIC_EXECUTION]: The orchestration scripts dynamically construct shell commands to invoke the Codex CLI. They use robust escaping techniques, such as Python's
shlex.quoteand POSIX-compliant single-quote wrapping in Bash, to ensure that user-supplied prompts do not lead to command injection. - [INDIRECT_PROMPT_INJECTION]: The skill provides an interface to an autonomous agent (Codex) that can perform file system modifications, which creates a potential surface for indirect prompt injection.
- Ingestion points: User-provided prompts in
codex-exec.sh,codex-goal.sh, andcodex-session.py. - Boundary markers: The skill uses
developer_instructionsto define subagent personas and constraints but does not wrap user prompts in explicit safety delimiters or include instructions to ignore embedded malicious data. - Capability inventory: Codex subagents are executed in the
workspace-writesandbox by default, allowing them to modify files and run project-specific commands. - Sanitization: The scripts do not perform semantic filtering or sanitization of the user-supplied prompts before delegation to the underlying agent.
Audit Metadata