codex-oracle
Warn
Audited by Gen Agent Trust Hub on Jun 23, 2026
Risk Level: MEDIUMCOMMAND_EXECUTIONDATA_EXFILTRATIONEXTERNAL_DOWNLOADS
Full Analysis
- [COMMAND_EXECUTION]: The skill instructs the agent to execute a local bash script located at
~/.claude/skills/codex-oracle/scripts/call-openai.sh. This script takes user-supplied input (thepromptparameter) as a command-line argument. This execution pattern presents a risk of command injection if the script does not properly sanitize or quote the input before processing it. - [DATA_EXFILTRATION]: The skill's primary purpose is to send local information, including "Specific code snippets or file contents," to OpenAI's API. While this is the stated functionality, it creates a significant data exposure surface where sensitive proprietary code or configuration details from the local environment are transmitted to a third-party service.
- [EXTERNAL_DOWNLOADS]: The skill involves network operations to communicate with OpenAI's API infrastructure for processing requests.
- [PROMPT_INJECTION]: The skill takes a raw text prompt as input to be forwarded to an external model. This is a surface for indirect prompt injection, especially when the agent is instructed to include local files or code in the prompt that may contain adversarial instructions targeting the external model's reasoning.
Audit Metadata