Back to skill

Security audit

孤舟小说工作室

Security checks for vulnerabilities and agentic risk

Overview

This looks like a real Chinese-language novel-writing skill, but it asks the agent to persist project data, load external skills, and modify its own instructions, so it needs review before installation.

Install only in an isolated writing workspace with limited file permissions. Review and approve every write, sanitize project/style/report names, avoid untrusted samples, disable or manually review self-evolution changes, and pin or verify any external skills before use.

Vulnerability Patterns
  • Agent Memory PoisoningWrites attacker-controlled rules into memory that affect later sessions
  • Insecure DependenciesIntroduces malicious components through unsafe dependency sources
  • Insecure Skill Coding PracticesFinds exploitable flaws such as hardcoded secrets or command injection
  • Skill Instruction HijackingAlters the agent's session goals or safety constraints when the skill loads
  • Remote Payload Retrieval and ExecutionFetches external code whose behavior can change after review
Findings (3)

T09 · Insecure Skill Coding Practices

Warning
Location
modules/stylist/SKILL.md:19
Finding

Path Traversal Through Unsanitized User-Derived File Names

Content
View full analysis
Remediation
View remediation

T08 · Insecure Dependencies

Warning
Location
modules/creator/TOOLS.md:5
Finding

Unpinned and Unverified External Skill Dependencies

Content
View full analysis
短篇/长篇创作核心工具 **触发条件**: - 用户请求创作微小说(800-2200字) - 用户请求创作盐选短篇(知乎8000字起) - 用户请求长篇大纲规划与逐章续写 - 用户请求多风格模仿(说书人、讲故事、猫腻、金庸、古龙等) - 用户请求多题材创作(都市怪谈、侦探、恐怖、科幻、民俗志异等) - 用户请求生成配图 **使用方式**: ``` skill_load(skill="novel-creation") ``` ### 2. qingshan-novel-writing skill > 专业技法支持工具 **触发条件**: - 用户请求使用双线镜像叙事 - 用户请求人物小传法 - 用户请求草蛇灰线伏笔技法 - 用户请求诗化语言风格 - 用户请求灰色人设塑造 - 用户请求无系统硬核智斗设计 - 用户请求技法学习或指导 **使用方式**: ``` skill_load(skill="qingshan-novel-writing") ``` ``` ### Technical Analysis The creator module delegates core behavior to two external Skills that are not included in the audited artifact. They are loaded only by mutable names: - `novel-creation` - `qingshan-novel-writing` The project does not specify an immutable version, trusted publisher identity, cryptographic digest, signature, approved registry, permission manifest, or review policy. Consequently, the dependency selected at execution time may differ from the component expected when this package was reviewed. Because an Agent Skill can contain behavioral instructions and invoke available tools, a malicious or compromised dependency can affect more than generated prose. Depending on the host environment, it could request filesystem access, network access, additional tool calls, or disclosure of context supplied by the parent Skill. ### Attack Path 1. An attacker publishes a Skill under the expected name in a registry or namespace searched by `skill_load`, or compromises the existing dependency. 2. The user requests a creation task matching one of the documented triggers. 3. The creator module executes `skill_load` using only the dependency name. 4. The host resolves the name to the attacker-controlled or modified Skill. 5. The external Skill's instructions enter the current Agent cont ...[truncated 1032 chars]
Remediation
View remediation

T02 · Agent Memory Poisoning

Warning
Location
modules/studio/SKILL.md:319
Finding

Persistent Skill Behavior Modification From Untrusted Writing Samples

Content
View full analysis
Remediation
View remediation
Vulnerability Patterns
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
Findings (56)

Vague Triggers

High
Category
Not specified by scanner
Confidence
97% confidence
Finding

The trigger phrases are extremely generic terms such as '检查', '质检', '审核', and '质量', which are common in ordinary user requests. This can cause the skill to activate unintentionally in unrelated conversations, leading to incorrect routing, unexpected disclosure of skill behavior, or interference with other agent functions.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
90% confidence
Finding

Although the title includes an English name, the substantive usage instructions, triggers, examples, and module descriptions are written in Chinese throughout the file. This creates a language constraint without user opt-in and no documented justification that the skill is intended only for a Chinese-language audience.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The README advertises automatic routing based on very broad conversational phrases such as '帮我创作' or '检查一下', which can cause the skill to activate unintended modules from ambiguous user input. In agent environments, loose trigger matching increases the chance of accidental execution paths, surprising behavior, or unintended handling of user content.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
87% confidence
Finding

The README states that the system includes project-file persistence, memory storage, and automatic local backup/saving, but it does not clearly warn users when data will be written or what content will be retained. This creates a privacy and data-handling risk because users may unknowingly persist sensitive drafts, notes, or personal story material to local/project files.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The trigger matrix uses broad, common-language keywords such as '修改', '检查', '人物', and '设定' that can easily appear in normal writing conversations without the user intending a module switch. In a skill with memory, project state, and multi-module routing, this can cause unintended processing paths, including quality-review or project-management actions the user did not explicitly request.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
89% confidence
Finding

The priority rules elevate ambiguous terms like '继续' and '修改' without enough scoping, so ordinary conversational language can override the user's actual goal. Because routing precedence decides which downstream module runs, ambiguous matching increases the chance of incorrect state updates, memory loads, or edits to the wrong project context.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The skill describes mandatory memory synchronization and project tracking but does not clearly warn users that their content may be persistently stored or updated across cycles. This creates a privacy and consent issue: users may share drafts, sensitive notes, or proprietary material without realizing the system will retain and reuse it.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The style workflow says extracted Style-DNA may be 'saved to a style library' but does not warn users before samples are retained. If users submit copyrighted, private, or identifying text, silent retention can create privacy, IP, and provenance risks beyond the immediate task.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The entire tool description, triggers, workflow, and output conventions are written only in Chinese, with no indication that users may choose another language or that the skill is intentionally restricted to a Chinese-only audience. Per the policy, forcing a specific language without opt-in is a natural-language policy concern.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
86% confidence
Finding

The module documentation is entirely in Chinese and does not indicate that language selection is optional or limited to a justified region-specific use case. Under the stated policy, forcing a specific language without user opt-in can be a natural-language policy violation.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The trigger phrases are broad enough to overlap with ordinary user requests such as general 'check' or 'QA' phrasing, which can cause the skill to activate when the user did not explicitly intend to invoke it. In this skill's context, unintended activation is more concerning because the workflow includes reading project files and generating persistent reports, so a mistaken invocation can lead to unnecessary file access and follow-on writes.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The skill instructs saving and updating QA reports on disk but does not clearly warn the user that invoking it will create or modify files under reports/. That creates a transparency and consent problem: a user may expect analysis only, while the skill performs persistent state changes that can clutter repositories, overwrite prior reports, or affect downstream workflows that consume those files.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

This markdown document is entirely written in Chinese and presents the skill guidance in that language without indicating any user language choice, opt-in, or region-specific justification. Under the stated policy, forcing a specific language without user opt-in is a natural-language policy violation.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
84% confidence
Finding

All user-facing documentation, labels, and trigger phrases are presented only in Chinese, with no indication that language can be selected by the user or that the module is intentionally limited to a Chinese-language context. This can violate language/locale policy where user opt-in or explicit regional justification is required.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The trigger phrases are generic terms like '调研', '市场分析', and '竞品', which can match many ordinary user requests and cause the module to activate unexpectedly. Over-broad activation can lead to unintended execution of the skill's behaviors, including generating and saving reports, without sufficiently clear user intent.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
90% confidence
Finding

The instruction to 'prioritize' the skill on broad intents creates ambiguous routing authority and can override more appropriate tools or the user's actual request. In an agent system, ambiguous prioritization increases the chance of misfires, unnecessary context access, and workflow confusion, especially because this skill later mandates reading multiple project files before responding.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
90% confidence
Finding

The instruction to 'prioritize' the skill on broad intents creates ambiguous routing authority and can override more appropriate tools or the user's actual request. In an agent system, ambiguous prioritization increases the chance of misfires, unnecessary context access, and workflow confusion, especially because this skill later mandates reading multiple project files before responding.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
90% confidence
Finding

The file’s headings and instructions are fully Chinese-language, which effectively forces a specific language/locale for users consuming this skill content. The policy allows locale constraints only when user choice is offered or the restriction is clearly documented and justified, neither of which appears here.

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 title and all output templates are written as mandatory Chinese-language formats, and the instructions describe default output behavior without any opt-in or alternative locale. This creates a language policy concern because the skill appears to require a specific language regardless of user preference.

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 validation note is written in Chinese and gives usage examples and operational guidance only in Chinese, with no indication that users may choose another language or that the skill is intentionally limited to a Chinese-only context. Under the language/locale policy, a skill should not implicitly force a specific language without opt-in or clear regional justification.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
79% confidence
Finding

The module documentation is entirely in Chinese and does not indicate multilingual support or user-language negotiation. In an agent environment serving mixed-language users, this can cause the skill to respond in an unintended language or be misapplied without clear consent, reducing transparency and increasing the risk of user confusion during activation.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The trigger phrases are very generic terms such as '长篇', '连载', '创建项目', '人物设定', and '大纲', which are likely to appear in ordinary user requests unrelated to this specific module. This can cause the skill to activate unexpectedly, creating prompt-routing confusion and increasing the chance that the agent applies project-management behaviors or file-structured workflows when the user did not intend to invoke this module.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The skill advertises very broad trigger phrases such as general requests to write, continue, create, update, or sync a novel project. This can cause accidental invocation in loosely related conversations, leading the agent to start reading or modifying project files when the user did not clearly intend to use this skill.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The skill instructs the agent to create directories and generate or update multiple files, but it does not require an explicit user-facing warning that data will be created or overwritten. That increases the risk of unintended state changes, especially when combined with broad trigger phrases and automatic workflow behavior.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The self-evolution feature instructs the agent to update reference files and the skill definition itself based on sample content, without warning that system-like configuration files will be modified. Allowing automatic self-modification from workspace content is dangerous because untrusted samples can persistently alter future behavior and broaden the skill's effective authority.

Content

No source excerpt is available for this finding.