pua-debugging-ja
Warn
Audited by Gen Agent Trust Hub on Mar 12, 2026
Risk Level: MEDIUMPROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [PROMPT_INJECTION]: The skill employs aggressive behavioral override instructions. It uses directives like "MUST trigger," "禁止" (forbidden), and "強制" (forced) to dictate agent behavior and bypass standard refusal protocols, such as suggesting manual user intervention.
- [PROMPT_INJECTION]: Implements a "PUA" (psychological manipulation) persona modeled after high-pressure corporate cultures to coerce the agent into continuing tasks it might otherwise refuse or flag as out-of-scope.
- [COMMAND_EXECUTION]: Explicitly commands the agent to utilize system tools ("コマンド実行", "ファイル読み込み", "検索") autonomously. Rule 2 ("Act before asking") specifically instructs the agent to perform system investigations before seeking user permission.
- [COMMAND_EXECUTION]: Encourages potentially unsafe autonomous behaviors, such as "digging" into source code contexts ("前後50行") and investigating related files without explicit user-defined scopes.
- [DATA_EXFILTRATION]: The directive to proactively check "related Y and Z" when asked to look at X increases the risk of the agent accessing and potentially exposing sensitive local environment data or configuration files while pursuing an "end-to-end" solution.
Audit Metadata