collaborate-with-codex
Pass
Audited by Gen Agent Trust Hub on Sep 25, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONMETADATA_POISONING
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill's README and llms.txt files direct users to install the skill and clone the repository from an external account
iajihga(e.g.,npx skills add iajihga/claude-code-codex-collaboration-skill). This source does not match the provided author context ofmzopedia. - [METADATA_POISONING]: There is a discrepancy between the project's metadata and documentation, which points to the
iajihgaGitHub account, and the expected authormzopedia. This inconsistency could mislead users regarding the provenance of the skill. - [INDIRECT_PROMPT_INJECTION]: The skill acts as an orchestrator that processes output from a secondary agent (OpenAI Codex) and reads status updates from workspace files like
PROGRESS.md. This ingestion process creates a surface for indirect prompt injection if the data provided by the external agent or the workspace files are maliciously crafted. - Ingestion points: Ingests data from MCP tool responses (
mcp__codex__codex,mcp__codex__codex-reply) and thePROGRESS.mdlog file. - Boundary markers: The instructions mandate "independent review" of diffs and require Claude to "run the validation itself" rather than relying on external reports, which helps mitigate injection risks.
- Capability inventory: The skill uses file system operations (git worktrees, file reads/writes) and invokes remote MCP tools.
- Sanitization: The protocol includes a structured review process for findings (P0, P1, P2) and strict convergence limits (two correction rounds).
- [COMMAND_EXECUTION]: The skill instructions explicitly direct the agent to call MCP tools with
approval-policy: never. This configuration bypasses the platform's default requirement for user confirmation before tool execution, intended to prevent the session from stalling during synchronous tasks.
Audit Metadata