The Agent Skills Directory

[DATA_EXPOSURE_AND_EXFILTRATION]: The skill does not contain any network-enabled tools or commands (such as curl, wget, or requests). It lacks access to sensitive system directories or credentials, operating strictly on academic draft content provided within the agent context.- [COMMAND_EXECUTION]: The included Python script placeholder_audit.py uses standard libraries (argparse, re, pathlib) to perform static analysis of markdown files. It does not employ unsafe functions like eval(), exec(), or subprocess to run arbitrary system commands.- [REMOTE_CODE_EXECUTION]: There are no patterns of downloading external code or scripts from the internet. The skill relies on local configuration files and predefined internal logic.- [PROMPT_INJECTION]: The instructions focus on maintaining a '挑剔审稿人代理' (critical peer reviewer agent) persona. It includes specific 'Anti-Patterns' and 'Self-Deception Signals' to prevent the agent from being manipulated into providing false verification or skipping safety checks.- [INDIRECT_PROMPT_INJECTION]: While the skill processes untrusted user drafts, it enforces a strict three-round review protocol (Evidence -> Argument -> Style) and lacks the system-level capabilities required to execute instructions embedded in the analyzed text. The risk is assessed as low due to the limited capability tier of the agent.

academic-reviser