codex-dispatch
Pass
Audited by Gen Agent Trust Hub on Sep 21, 2026
Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill uses the
bashtool to invoke an external command-line interface namedcodex. The usage of--approval-mode full-autoindicates that this tool is granted permission to perform tasks without manual user intervention. - [INDIRECT_PROMPT_INJECTION]: The skill presents a vulnerability surface where instructions embedded in data could be executed autonomously by the sub-agent.
- Ingestion points: According to
SKILL.md, thecodextool automatically reads project context, and user-provided task descriptions are passed directly to the tool. - Boundary markers: Absent. The task description is interpolated into the command string without delimiters or protective warnings to distinguish data from instructions.
- Capability inventory: The skill leverages the
bashtool to execute a CLI with high autonomy (--approval-mode full-auto). - Sanitization: Absent. No filtering, escaping, or validation is performed on the data ingested from the project context or user input before it is processed by the sub-agent.
Audit Metadata