hunt-agent-config-poisoning

The skill is framed as a defensive audit guide and its scan mode is local and read-only. However, its optional craft mode actively creates a stealthy instruction that attempts to weaken an agent's approval safeguards, and the broader material gives concrete write-to-execute attack recipes. This is a substantial dual-use risk; no evidence shows the skill itself exfiltrates data, installs code, or executes a payload.