run-research-and-save-files-by-codex
Fail
Audited by Gen Agent Trust Hub on May 19, 2026
Risk Level: HIGHCOMMAND_EXECUTIONREMOTE_CODE_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
- [COMMAND_EXECUTION]: The skill invokes the
codex execcommand to perform research tasks. Technical documentation inreferences/codex-exec-contract.mdspecifies the use of the--dangerously-bypass-approvals-and-sandboxflag. This flag is designed to disable security approvals and execution sandboxing, which removes critical safety barriers for the commands generated and run by the AI. - [REMOTE_CODE_EXECUTION]: The core mechanism of the skill involves the orchestrator (Claude) generating prompts that are subsequently executed by a subprocess (
codex). This effectively creates a remote code execution pathway where AI-generated instructions are carried out on the host system. Thereferences/codex-exec-contract.mdfile demonstrates this with a smoke test that executes arbitrary filesystem and text processing tasks defined in the prompt. - [EXTERNAL_DOWNLOADS]: The skill is designed to perform extensive web research, which necessitates downloading content from untrusted external domains. This content is then synthesized and processed. This creates a surface for indirect prompt injection, where malicious instructions hidden in web pages could influence the behavior of the research subprocesses or the orchestrating agent.
- [SAFE]: The installation instructions in
INSTALL.mduse the standardnpxpackage manager to add the skill from a known repository path (yigitkonur/skills-by-yigitkonur).
Recommendations
- AI detected serious security threats
Audit Metadata