Back to skill

Security audit

一盏神灯 Yizhan Shendeng

Security checks for vulnerabilities and agentic risk

Overview

The skill’s main purpose is coherent, but it asks agents to run long automated local/cloud workflows and touches account/network features that should be reviewed before installation.

Install only if you are comfortable with a long-running, file-writing workflow that may use sub-agents, local OCR/scripts, paid-model capacity, and account-balance checks. Before running it, confirm where it may write files, whether balance/API checks are allowed, whether cross-channel notifications are desired, and whether optional installation or book-library archival steps should be disabled.

Vulnerability Patterns
  • Skill Instruction HijackingAlters the agent's session goals or safety constraints when the skill loads
  • Agent Memory PoisoningWrites attacker-controlled rules into memory that affect later sessions
  • Remote Payload Retrieval and ExecutionFetches external code whose behavior can change after review
  • Embedded Malicious CodeShips malicious scripts inside the skill and executes them locally
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs
Vulnerability Patterns
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • MCP Least PrivilegeUnderdeclared Capability, Wildcard Permission, Missing Permission Declaration
Findings (25)

Tp4

High
Category
MCP Tool Poisoning
Confidence
95% confidence
Finding

The declared description says the skill distills long-form material into skill packages, but the body also performs internal QA, file-system workflow management, quote verification, and pipeline auditing behaviors not surfaced to the user. This mismatch weakens informed consent and can cause the skill to perform broader actions than a user reasonably expects from the description alone.

Content

No source excerpt is available for this finding.

Ae1

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 421)May include surrounding context.

md
逐个打开每个 `SKILL.md`,检查四件事:

Ae1

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 450)May include surrounding context.

md
逐个打开每个 `SKILL.md`,检查四件事:

Ae1

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 464)May include surrounding context.

md
逐个打开每个 `SKILL.md`,检查四件事:

Undeclared Tool Scope

Medium
Category
MCP Least Privilege
Confidence
88% confidence
Finding

The skill clearly instructs file-system reads, workspace inspection, and source-material processing, but it does not declare any explicit tool scope or allowed-tools boundary. That creates an authorization ambiguity where a host may permit broader file access than users expect, increasing the chance of unintended access to local files during execution.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
89% confidence
Finding

The skill mandates fully automatic execution and automatic continuation across queued files, including creation of multiple workspace files and long-running processing. Without a prominent upfront warning and consent boundary, users may not realize they are authorizing sustained local processing and cascading actions across multiple inputs.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
98% confidence
Finding

The instruction '全程用中文交流' mandates a single language for the entire skill flow and does not offer the user a choice or opt-in. This is a natural-language locale policy issue because it forces a language preference regardless of user needs or environment.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

The workflow goes beyond distillation by instructing cross-channel messaging and provider account-balance checks. Those actions touch external systems and user account metadata unrelated to the core task, expanding the skill's effective authority and data exposure surface.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
90% confidence
Finding

The skill directs the agent to inspect provider balances and local system resources such as free memory. These checks are operationally useful, but they are not essential to the user-facing purpose and unnecessarily expand access to sensitive environment and account information.

Content

No source excerpt is available for this finding.

External Transmission

Medium
Category
Data Exfiltration
Confidence
94% confidence
Finding

The skill contains an explicit external API endpoint for balance checks and also discusses cross-channel communications. Even if intended as read-only accounting, outbound network access creates a real data-transmission path and can expose account metadata, timing information, and potentially document-related operational context outside the local task boundary.

Content

Scanner excerpt · SKILL.md (reported line 20)May include surrounding context.

md
- 每一步开始和结束向用户播报一行进度,**带上工作目录路径**
- 全程用中文交流。原文是外文时,原文摘录保留外文原句,其余一律中文
- 子 Agent 跑在付费模型上:重要批次前确认账户额度,长任务期间留意余额。额度将尽时按"核心优先"裁剪(先保提取/验证/构造,测试转精简版);额度耗尽会令在跑的子 Agent 集体中断,比省着跑更贵。备选模型渠道写好切换方式,只作救场。用户问"蒸一本要多少钱"的估法:①用 pymupdf 采样出源文件页数与字数;②数子代理量(提取=批数×5、验证=包数、构造=技能数);③拿上一本全流程用 `hermes insights --days 1` 实测做基准折算(经验值:70 万字级书全程 ≈ $1–1.5),报区间不报点数
- **成本口径(记真账)**:`hermes insights` 对没配定价的模型会低报(大批会话记成 Unknown),只能当参考;真实消耗=开工前与收尾各查一次账户余额、把「余额+时间」快照追加进日志文件,用前后差值核算。余额查询走只读接口(DeepSeek:`GET https://api.deepseek.com/user/balance`),key 在脚本进程内读取、绝不打印进对话。
- **负载纪律(硬约束)**:本地重活与云端批次不要叠加——对并行子代理批次来说,本地再跑 OCR、大编译这类重任务是纯干扰;用户报告"别的任务跑不动"时,先暂停或停掉本地重活(可断点续跑的进度保留、随时可续),云端批次照常推进。本地多进程任务一律 ≤2 进程(扫描版 OCR 的实测安全值;多进程重活曾拖死用户整机);启动重活前先查空闲内存(`head -3 /proc/meminfo`,在 Windows/MSYS 下即系统内存),低于 3GB 不再加新进程
- **无人值守过夜模式**:用户说"我要睡了/都来吧/继续跑"授权夜跑时,切最保守配置,守住一条底线——**任何一步出问题都必须能自己兜底,不允许出现需要用户半夜爬起来修的操作**:只做可断点续跑的任务;不升级、不重装、不动系统配置、不做不可恢复的批量操作;多本排队时 OCR 串行(一次只开一个 OCR 任务);一本完成自动接队列下一本;早上给"完成了什么+还剩什么"的汇报
- 内容包含违法操作或伤害他人的具体方法,立即停止并说明原因

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The skill instructs sending completion notifications to user-specified external channels via a messaging command. That creates an outbound communication path that can leak document titles, workspace paths, progress state, or other metadata outside the current session without being part of the core distillation function.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
84% confidence
Finding

The OCR path notes that local conversion may be slow, but the skill lacks a single high-level warning that OCR and background processing can consume significant CPU, memory, and disk over long periods. In an unattended setting, that can degrade the user's machine or interfere with unrelated workloads.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
87% confidence
Finding

Although the workflow says not to auto-install by default, it still offers direct installation on request and also performs optional archival moves in a maintained book library. Those actions modify user configuration and file organization beyond merely producing a skill package, creating integrity risk if paths or target locations are wrong.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
78% confidence
Finding

This markdown file documents a workflow that writes back to existing files using overwrite mode and repeatedly edits production candidate files. Although there is one backup note for a specific recovery step, the playbook lacks a clear user-facing warning near the main write path that these operations can change existing data and should be treated as potentially destructive.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
88% confidence
Finding

The guidance explicitly accepts overlapping trigger concepts across sibling skills and only suggests cleanup as optional, which can cause ambiguous or overly broad activation. In an agentic system, this increases the chance that the wrong skill is invoked for a user request, leading to boundary violations, unintended file/workflow actions, or unreliable task routing.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The entire skill specification is written to require responses in Chinese and fixed Chinese output formats such as '## 案例:' and '第 1 块:找到 2 条'. This imposes a specific language/locale behavior without documenting user choice or opt-in, which matches the language-policy violation category.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The entire skill file is written as an instruction set in Chinese and does not indicate that users may choose another language or locale. Under the policy, forcing a specific language without opt-in is a natural-language policy violation unless the locale restriction is explicitly justified.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The file's operational instructions are entirely in Chinese and do not provide any user opt-in, alternative language, or justification for restricting use to that language. This can violate language/locale policy because it implicitly forces a specific language for using the skill.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The file mandates Chinese-language interaction throughout the skill content, but does not provide user opt-in or an alternative language option. Under the policy, forcing a specific language without user choice is a natural-language policy violation unless the locale constraint is explicitly justified.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The file mandates operation in Chinese throughout the skill instructions and does not provide any user opt-in, fallback, or justification for a Chinese-only locale. This can violate language/locale policy when users are not given a choice of language.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

This markdown template is entirely written as an instruction set in Chinese and directs the user to produce 骨架.md using that template. There is no indication that Chinese is optional, nor any justification that the skill is region- or locale-specific, which makes it a natural-language locale policy concern.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

The template explicitly requires a Chinese skill name in the main heading, and the surrounding document is written as prescriptive guidance for all skills. This imposes a specific language/locale requirement without indicating that users may choose another language or that the constraint is limited to a justified region-specific context.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
98% confidence
Finding

This markdown template forces a specific language/locale for all users of the skill artifact, and there is no opt-in, alternative language option, or justification that the skill is intended only for a Chinese-speaking or region-specific context. That matches the natural-language policy violation criterion for language or locale restrictions.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

The instruction '全程中文' forces a specific language for the task. This is a natural-language locale/language policy concern because the file does not present language choice or explain a justified region-specific constraint.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

This code file contains natural-language instructions and status messages entirely in Chinese, including the top-level usage text and runtime output strings. Because the file provides no indication of language choice or opt-in, it enforces a specific language/locale in a way that can violate language-choice policy.

Content

No source excerpt is available for this finding.