The Agent Skills Directory

[METADATA_POISONING]: The skill makes several deceptive claims regarding AI models and performance metrics.
It references fictitious models such as "GPT-5.2" and "GPT-5.1-thinking" which do not exist in the referenced OpenAI ecosystem.
It cites specific but unverifiable benchmarks (e.g., "76.3% SWE-bench") for these non-existent models to build false trust.
[COMMAND_EXECUTION]: The skill provides instructions that create a command injection vulnerability.
It directs the agent to use syntax like echo "your prompt here" | codex exec .... If the user input contains shell metacharacters (e.g., backticks or command substitutions), it could lead to arbitrary command execution on the host system.
[METADATA_POISONING]: The skill explicitly instructs the agent to suppress error output by default.
Using 2>/dev/null for all commands hides stderr, which is a concealment tactic. This prevents the user from seeing potential warnings, errors, or evidence of malicious activity generated by the underlying codex CLI tool.
[PRIVILEGE_ESCALATION]: The skill encourages the use of high-risk flags.
It promotes the use of --sandbox danger-full-access and --full-auto, which grants the CLI tool broader system permissions and autonomous execution capabilities, increasing the potential impact of a compromise.

codex