The Agent Skills Directory

[PROMPT_INJECTION]: The skill uses direct interpolation of user input via $ARGUMENTS without any boundary markers or sanitization, making it vulnerable to injection attacks that could redefine the execution plan.
[PROMPT_INJECTION]: Instructions explicitly mandate bypassing human-in-the-loop safety measures, stating "Do not ask for plan approval" and "Complete automation ... without any human intervention."
[COMMAND_EXECUTION]: The skill grants full Bash tool access to all sub-agents and encourages autonomous execution, which can lead to arbitrary command execution on the host system.
[DATA_EXFILTRATION]: By granting full Read and Bash capabilities to autonomous sub-agents while processing untrusted input, the skill creates a significant risk of local file exposure or exfiltration if a sub-agent is manipulated.
[PROMPT_INJECTION]: The skill uses directive language like "严禁" (strictly prohibit) and "must allow" to override default agent behaviors and safety constraints regarding tool usage and reporting.

do-agent