Security audit
Security checks for vulnerabilities and agentic risk
This is a markdown-only intent-classification and routing guide with broad matching examples, but it does not install code, request permissions, persist data, or execute actions itself.
Installers should understand that this skill may classify common Chinese phrases like polite requests as task intent. Use its output as advisory routing only, and require explicit user confirmation before invoking any downstream skill that can edit files, run commands, search private data, or store memories.
The activation phrases "识别意图 / 分类任务 / 智能路由" are broad, generic terms that can naturally appear in ordinary conversation or user requests. In an agent environment, this can cause unintended invocation of the skill and accidental routing into downstream skills, expanding the chance of unauthorized actions or prompt/flow hijacking through misclassification.
The task-intent keywords include very common words like "帮我", "请", and "做", which are ubiquitous in normal user speech. This makes the classifier over-match benign conversation as actionable task intent, increasing the risk that the agent routes users into execution-oriented workflows without sufficiently clear intent.
The keyword-based classifier relies on substring matches for extremely broad tokens, including common Chinese function words and basic action verbs. Because these checks lack contextual guards, boundary checks, or disambiguation, they can misclassify routine text and trigger incorrect routing, especially to skills associated with code generation, learning, or task automation.
The regex pattern ^帮我.* matches any input starting with a very common phrase, so ordinary polite requests will be treated as task-routing triggers. In context, this is more dangerous because the skill is explicitly designed as a dispatcher to other skills, meaning broad matching can fan out into unintended downstream capabilities.
The skill metadata and body are written as Chinese-only user-facing instructions and examples, with no indication that users may interact in other languages or opt into Chinese. This can violate language/locale policy when a skill imposes a language without user choice.
No suspicious patterns detected.
