The Agent Skills Directory

[PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection as it incorporates untrusted user input into a prompt for the Codex AI model.
Ingestion Point: The query argument from $ARGUMENTS is directly used to construct the prompt in SKILL.md.
Boundary Markers: The skill does not use delimiters or specific instructions to isolate user input from system prompts.
Capability Inventory: The run-codex-exec.sh script executes the codex exec command, which can perform complex analysis based on its input.
Sanitization: There is no evidence of input validation or escaping for the user-provided query.
[EXTERNAL_DOWNLOADS]: The scripts/check-codex.sh script references external installation sources for the Codex CLI.
Evidence: The script suggests installing @openai/codex via npm or the openai/tap/codex Homebrew tap.
Source: These references point to OpenAI, which is a recognized and trusted organization. The documentation is provided neutrally as setup instructions.
[COMMAND_EXECUTION]: The skill executes local shell scripts and a binary CLI tool.
Evidence: It runs scripts/check-codex.sh and scripts/run-codex-exec.sh, which in turn executes the codex binary.
Security: The skill uses a HEREDOC and temporary files to pass input to the binary, which is a safe practice that avoids shell command injection vulnerabilities.

oracle-codex