The Agent Skills Directory

Privilege Escalation (HIGH): The skill encourages the use of the --sandbox danger-full-access flag, which grants the tool unrestricted access to the host system and network, bypassing typical security boundaries.
Metadata Poisoning (MEDIUM): The skill uses deceptive and fictional claims about 'GPT-5.2' and specific benchmark stats (76.3% SWE-bench) to establish false authority and persuade the agent to utilize the high-privilege tool.
Obfuscation (MEDIUM): Instructions to append 2>/dev/null to all execution commands serve to hide stderr output. While described as suppressing 'thinking tokens,' this effectively silences error messages, warnings, or logs that would otherwise alert a user to malicious or failing behavior.
Command Execution (LOW): The skill relies on piping user input via echo "prompt" | codex exec, which presents a shell injection risk if the agent interpolates untrusted external data into the prompt string without adequate sanitization or escaping.
Prompt Injection (LOW): The instruction to always use --skip-git-repo-check bypasses a built-in safety mechanism designed to verify the integrity or context of the code being modified.

codex