Ae1
- Category
- analysis-evasion
- Confidence
- 100% confidence
- Finding
Referenced artifact was not completely inspected
- Content
md - `scripts/night_light_tail.sh`
Security audit
Security checks for vulnerabilities and agentic risk
This skill’s night-light automation is coherent, but it needs Review because it can control home devices from broad phrases and stores sensitive nighttime routine preferences without clear consent or retention limits.
Install only if you are comfortable with an assistant controlling Xiaodu-connected lights/scenes from short Chinese phrases and saving night-use preferences. Before using it in a shared home, set clear activation wording, require confirmation for ambiguous phrases, and decide whether XIAODU_CONTEXT.md or MEMORY.md may store nighttime routine details.
Referenced artifact was not completely inspected
- `scripts/night_light_tail.sh`
Referenced artifact was not completely inspected
- `scripts/night_light_tail.sh`
The trigger list includes very broad phrases such as “太黑了” and “我要喝水,开一下灯”, which are common everyday utterances that may not reliably indicate consent for this specific night-assist automation. In a home-control skill, overbroad activation can cause unintended device actions at night, including turning on the wrong lights or starting an automation sequence without sufficiently clear intent.
The skill instructs the agent to persist user preferences about nightly movement patterns, room/device choices, and behavior defaults into memory files. In this context, those preferences can reveal sensitive household routines and inferred health or age-related information, and the skill provides no retention limits, minimization rules, consent checks, or access controls.
All user-facing inputs and expected phrasing in the test cases are written exclusively in Chinese, including the expected user-facing response style, with no indication that language choice is optional or that the skill is intentionally limited to a Chinese-speaking locale. This can violate language/locale policy when a skill implicitly forces one language without user opt-in.
The test cases use broad natural-language inputs such as “太黑了” that can plausibly occur in ordinary conversation, yet the file does not document stronger activation constraints, confirmation requirements, or contextual gating. In a home automation skill that controls nighttime lighting and possibly temperature, this can cause unintended execution from ambient speech, misrecognition, or overlapping intents.
The examples around minimal confirmation explicitly allow direct execution for common phrases like “开一下走廊灯,” which may overlap with casual speech directed at another person or picked up from media. Because the skill performs physical-world actions, reducing confirmation too aggressively increases the chance of unintended device control.
The trigger list includes broad everyday utterances such as '太黑了' and '我要喝水,开一下灯', which can plausibly occur outside the intended elder night-assist scenario. In a home-control skill, unintended activation can cause device actions like lights turning on unexpectedly at night, creating privacy, safety, and usability risks even without a direct code exploit.
The notes direct the system to persist behavioral preferences such as nighttime routines and lighting habits into XIAODU_CONTEXT.md without any mention of user notice, consent, retention limits, or access controls. Because this skill targets elderly nighttime behavior, the stored data can reveal sensitive daily patterns and occupancy habits, increasing privacy and profiling risk if over-collected or exposed.
The description says to use the skill when the user says phrases like the listed examples, but does not clearly define when similar requests should instead route to a normal lighting skill. This leaves the activation boundary ambiguous, especially because the skill itself says its goal is not ordinary light control.