Security audit

得到大脑会员式思考技能包

Security checks for vulnerabilities and agentic risk

Overview

This is a markdown-only Chinese writing and thinking skill that processes user-provided text, with the main caution that file paths should only point to content the user intends to share.

Install only if you want a Chinese-language writing and thinking workflow. When using review or creation modes, paste the material directly or provide paths only to files you are comfortable having the assistant read; do not point it at credentials, private system files, or unrelated personal archives.

SkillSpector

Vulnerability Patterns

Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code

Findings (6)

Vague Triggers

Medium

Confidence: 89% confidence
Finding: The skill instructs the agent to auto-trigger whenever the user's intent appears to fit one of six broad categories, even without explicit user selection. This can cause over-application of the skill, override normal conversational handling, and increase the chance the agent follows skill-specific behavioral constraints in situations where the user did not actually request them.

Vague Triggers

Medium

Confidence: 86% confidence
Finding: The trigger table uses very common phrases such as '帮我看看', '写个文案', or '发散一下' without guardrails or exclusions, making accidental invocation likely during ordinary conversation. Broad everyday triggers can hijack unrelated requests into this skill's workflow and cause the agent to apply mode-specific rules, tone, or limitations in contexts where they are inappropriate.

Natural-Language Policy Violations

Medium

Confidence: 91% confidence
Finding: The skill content is entirely written to enforce Chinese-language behavior and provides no mechanism to honor the user's preferred language or request clarification. This can cause unintended language output, reduce accessibility, and create compliance or usability issues in multilingual environments, even though it is not an exploit-oriented security flaw.

Missing User Warnings

Medium

Confidence: 95% confidence
Finding: The skill explicitly allows users to provide arbitrary file paths for the agent to read, but it does not define any scope restrictions, safe directories, consent language, or checks against sensitive local/system paths. In an agent environment with filesystem access, this can enable unintended disclosure of secrets, configuration files, private notes, or other local data if a user is tricked into supplying a path or the model over-reads beyond the intended workspace.

Vague Triggers

Medium

Confidence: 86% confidence
Finding: The routing eval explicitly treats vague phrases like “帮我看看” as sufficient to trigger a mode selection, but it does not define strong disambiguation rules or hard exclusions. In a skill that auto-routes user intent across critique, rewriting, and content generation modes, this can cause the agent to infer the wrong task and produce unintended analysis or transformation, increasing prompt-injection and overreach risk.

Natural-Language Policy Violations

Medium

Confidence: 90% confidence
Finding: The skill content is written entirely in Chinese and specifies a fixed output style and tone without indicating that the assistant should adapt to the user's preferred language. This can cause the assistant to respond in an unintended locale, degrading usability, causing misunderstanding, and in some deployments bypassing product-level language expectations or accessibility requirements.

Static analysis

No suspicious patterns detected.