The Agent Skills Directory

[PROMPT_INJECTION]: The file references/score_program.md contains instructions that command the agent to "LOOP FOREVER" and "Never stop." It explicitly directs the agent to bypass human-in-the-loop oversight by not pausing for user confirmation during the optimization cycle, which overrides standard safety protocols.
[COMMAND_EXECUTION]: The script scripts/run_scoring.py executes an external CLI tool using subprocess.run. The path to this tool is derived from the CODEX_CLI_PATH environment variable, which could allow for the execution of untrusted binaries if the environment is misconfigured.
[PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection in scripts/run_scoring.py because it interpolates raw content from user-provided articles directly into the prompt without sanitization.
Ingestion points: Article files are read from the assets/articles/samples/ directory by scripts/run_scoring.py.
Boundary markers: The script uses basic markdown headers as delimiters but provides no instructions for the model to ignore potential instructions embedded within the article text.
Capability inventory: The skill has the capability to execute shell commands (codex) via the subprocess.run function.
Sanitization: No escaping, validation, or filtering is performed on the article content before it is added to the prompt.

score-optimizer