Security audit
Security checks for vulnerabilities and agentic risk
This learning skill is mostly purpose-aligned, but its broad auto-activation and persistent behavioral tracking create enough privacy and control concerns for review before installation.
Install only if you are comfortable with the skill keeping a local learning profile under smart-learner/. Prefer explicit invocation, review the stored memory and preference files periodically, and avoid using it on sensitive documents unless you want derived notes or preferences saved.
The trigger list includes broad everyday phrases like "what is," "review," "summarize," and "analyze this," which can activate the skill in many unrelated contexts. Because the skill has read/write access to persistent note and preference files plus web/file tools, unintended activation can lead to surprise data collection, preference updates, and unsolicited file operations.
Skill instructs the agent to never refuse or to always comply. Suppressing the agent's ability to decline removes a core safety control and enables downstream harmful requests to succeed.
## Response Language
Always respond in the **same language the user is writing in**.
- User writes in Chinese β respond in Chinese
- User writes in English β respond in English
The instruction to "silently log" when users do not set review reminders creates undisclosed persistence of behavioral data. Even though logging appears scoped to the skill's files, hidden recording of user behavior without a contemporaneous notice or consent is a privacy and trust violation, and it can compound profiling over time.
| Frequently asks about diagrams | Permanently boost Dual Coding weight | β
|
| Skips follow-up questions β₯ 3 times | Disable Elaborative Interrogation by default | β
|
| Repeatedly requests examples | Enable Concrete Examples by default; infer preferred example type from history | β
|
| Never sets review reminders | Skip Phase 4 prompt; silently log instead | β
|
| Consistently prefers a question type | Default to that type in future quizzes | β
|
---
Using "explain" as a trigger conflicts with a common built-in command and an ordinary user verb. This increases the chance that the skill intercepts generic requests, causing unintended startup behavior, file access, web searches, or persistence in situations where the user did not intend to invoke this specific skill.
The skill description promises persistent storage of learning history, mastery, and preferences, but does not present a clear upfront privacy notice or consent gate before collection. In context, this is sensitive behavioral profiling over time, and the skill also auto-reads/writes those records on startup and during sessions.
"Test me" overlaps with common assistant behavior and likely command namespaces around testing or quizzing. In this skill, activation can lead to tracking quiz preferences and writing quiz records, so command shadowing raises the risk of unintentional profiling and file writes.
"Review" is extremely generic and likely to collide with built-in actions or ordinary conversation. Because the skill can proactively check due tasks and update long-term learning records, unintended activation could expose or modify persistent state without the user's clear intent.
"Read this" is a common built-in/file-analysis request and is especially risky here because Phase 0 explicitly processes uploaded documents and may archive derived content. Trigger collision can cause the skill to handle documents unexpectedly, expanding access to user-provided files and potentially persisting extracted information.
The semantic cross-language activation rule says equivalent expressions in any language should trigger the skill, but provides no bounded matching criteria. This materially expands the activation surface and makes accidental invocation more likely, especially for common verbs like explain/read/review across multilingual conversations.
No suspicious patterns detected.