The Agent Skills Directory

[PROMPT_INJECTION]: The instructions employ adversarial language such as 'attack', 'destroy', and 'fatal weaknesses' strictly as metaphors for logical critique. The skill does not attempt to override platform safety filters or system instructions.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process user-provided ideas for review. While it lacks explicit delimiters for user content, the skill does not exhibit capabilities for file system access, network communication, or command execution, mitigating the potential impact of injection through data.
[DATA_EXFILTRATION]: Analysis confirmed the absence of network operations, hardcoded credentials, or access requests to sensitive environment variables or system configuration files.
[COMMAND_EXECUTION]: No evidence of shell command execution, subprocess invocation, or dynamic context injection was found in the instructions or the framework matrix.
[METADATA_POISONING]: Metadata fields including name and description were reviewed and found to be consistent with the skill's stated purpose, with no deceptive instructions detected.

thinking-opponent