Security audit
Security checks for vulnerabilities and agentic risk
This is a text-only skill that adds calming reminders during frustrated conversations, but some examples are broader and less scoped than the main description.
Install only if you want the assistant to occasionally add a brief grounding reminder when conversations become frustrated. Be aware that the examples should be tightened so the skill does not over-interpret general stress or imply memory-file writes.
The instruction "Detect the user's language and respond in kind" imposes a language-selection policy automatically rather than offering the user a choice. The policy also resolves mixed-language conversations by choosing the dominant language, which may override user preference without opt-in.
The example promises '我会主动写进记忆文件', implying persistent memory writes that are not disclosed in the skill metadata, which only describes appending a grounding reminder. This can mislead downstream agents into storing user data without clear authorization or expectation, creating privacy and scope-creep risks.
The 'Correct response' examples prescribe replies in Chinese, and the file does not indicate that language choice should follow the user's preference or be optional. This creates a natural-language policy concern because it effectively hard-codes a locale/language behavior without user opt-in or documented justification.
These examples expand the detector from its stated scope of frustration directed at the AI into broader emotional-distress inference, including a late-night heuristic and a statement like '我不行了' about project stress. That increases the chance of unsolicited mental-state assessment and intervention in situations unrelated to the product’s documented behavior, creating privacy, safety, and overreach risks.
This response example is written as a required 'Correct response' in Chinese, but the file provides no opt-in, language-selection guidance, or region-specific rationale. As written, it suggests the skill should force a specific language behavior across cases.
The file instructs that, if triggering, the skill should use a specific Chinese grounding line. Without any note to adapt to the user's language or an explicit locale limitation, this is a language policy violation under the stated rules.
