AI agent skills pass every scanner. 87% still degrade agent safety

have been looking into a skill repo that is "safe". If I want to drop one of these hardened skills into production, what's the failure mode? are you also exploring if the default guardrails ever block " legitimate" workflows?