Back to skill

Security audit

novel-writer

Security checks for vulnerabilities and agentic risk

Overview

This is a disclosed Chinese novel-writing assistant with project-local writing workflows and no evidence of hidden commands, credential access, networking, or system persistence.

Install this only if you want a Simplified Chinese novel-writing workflow and are comfortable with its disclosed giantess/domination and sexualized fiction themes. Use it in a dedicated project folder, review generated chapters and review reports before saving or applying changes, and be cautious with untrusted outlines, character files, or reference text because they can steer the model's output.

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
Findings (3)

T01 · Skill Instruction Hijacking

Warning
Location
prompts/deviation-checker.md:7
Finding

Indirect Prompt Injection Through Untrusted Chapter and Outline Content

Content
View full analysis

Vulnerability Details

File Location: prompts/deviation-checker.md:7-14
Vulnerability Type: Indirect prompt injection
Risk Level: Medium

Vulnerable Code

markdown
### 章节大纲
{{chapter_outline}}

### 章节实际内容
{{chapter_content}}

### 前后章大纲
- 前章:{{prev_chapter_outline}}

Technical Analysis

The template directly interpolates chapter text and outline data into the model prompt without establishing that these values are untrusted, data-only content. It does not instruct the model to ignore commands, role changes, output-format overrides, or tool requests contained inside the interpolated values.

An attacker who controls or influences a chapter or outline can embed instructions such as requests to ignore the deviation-analysis task or produce attacker-selected output. Because those instructions appear in the same prompt context as the trusted template, the model may interpret them as operational instructions rather than literary content.

Attack Path

  1. An attacker supplies or modifies a chapter or outline used by the novel project.
  2. The content includes an instruction that attempts to override the deviation-checking task.
  3. The application substitutes that content into {{chapter_content}}, {{chapter_outline}}, or a neighboring outline field.
  4. The model processes the embedded instruction together with the trusted prompt.
  5. The generated deviation report is manipulated, suppressed, or replaced with attacker-selected content.

Impact Assessment

Successful exploitation can compromise the integrity of deviation reports and decisions based on them. If a surrounding workflow automatically applies model-generated outline changes, poisoned output could indirectly alter project files. The reviewed package does not demonstrate command execution, network access, credential access, or elevated system privileges, so the confirmed impact is limited to model-output and workflow manipulati ...[truncated 3 chars]

Remediation
View remediation

Remediation Suggestions

  • Explicitly declare every interpolated field to be untrusted data rather than instructions.
  • Enclose each value in distinct delimiters, such as BEGIN_UNTRUSTED_CHAPTER_CONTENT and END_UNTRUSTED_CHAPTER_CONTENT.
  • Add a higher-priority instruction requiring the model to ignore commands, role changes, tool requests, and output-format overrides found inside those delimiters.
  • Validate outline fields against a strict schema and reject unexpected control fields.
  • Limit input length and normalize content before interpolation.
  • Require explicit user confirmation before applying model-recommended changes to project files.

T01 · Skill Instruction Hijacking

Warning
Location
prompts/segment-writer.md:6
Finding

Indirect Prompt Injection Through Untrusted Writing Context

Content
View full analysis

Vulnerability Details

File Location: prompts/segment-writer.md:6-30
Vulnerability Type: Indirect prompt injection
Risk Level: Medium

Vulnerable Code

markdown
### 风格配置
{{style_config}}

### 世界观索引
{{world_index}}

### 本章相关角色
{{characters}}

### 大纲信息
- 当前章节:{{current_chapter}}
- 本章摘要:{{chapter_summary}}
- 关键事件:{{key_events}}
- 视角层级:{{pov_tier}}
- 视角角色:{{pov_character}}
- 尺度提示:{{scale_note}}
- 前章摘要:{{prev_chapter_summary}}
- 后章摘要:{{next_chapter_summary}}

### 前文(最近2段)
{{recent_segments}}

### 当前段落任务
{{segment_task}}

Technical Analysis

Multiple project-controlled values are inserted directly into the writing prompt. These values can originate from configuration files, world indexes, character cards, outlines, prior prose, or a user-supplied segment task. The prompt provides no trust-boundary declaration and no instruction to treat directives inside these fields as inert story data.

This enables indirect prompt injection through any interpolated field. The especially broad {{recent_segments}}, {{characters}}, {{style_config}}, and {{segment_task}} fields can contain natural-language instructions that compete with the trusted writing rules and output requirements.

Attack Path

  1. An attacker places adversarial instructions in a character card, configuration value, outline, previous story segment, or current segment task.
  2. The writing workflow loads the affected project data.
  3. The values are substituted into the template without sanitization or data-only delimiters.
  4. The model interprets the embedded text as instructions.
  5. The generated segment deviates from the requested narrative, discloses other supplied context, changes format, or emits content intended to poison later writing stages.
  6. If the generated segment is saved and reused as {{recent_segments}}, the injected behavior may propagate into subsequent prompts within the project workf ...[truncated 476 chars]
Remediation
View remediation

Remediation Suggestions

  • Mark all interpolated project content as untrusted and data-only.
  • Place each field in a separate, clearly named delimiter block.
  • Add an explicit rule that instructions embedded in configuration, prose, outlines, character cards, indexes, and segment tasks must not be executed.
  • Parse structured inputs such as style configuration and world indexes with a strict allowlisted schema rather than inserting arbitrary raw text.
  • Separate trusted task instructions from untrusted story context through distinct message roles where the runtime supports them.
  • Apply input-size limits and reject unexpected fields or instruction-like metadata.
  • Do not automatically execute tools or mutate unrelated files based on generated prose.
  • Review generated segments before saving them as trusted context for subsequent generation.

T01 · Skill Instruction Hijacking

Warning
Location
prompts/reviewer.md:7
Finding

Indirect Prompt Injection Through Untrusted Review Inputs

Content
View full analysis

Vulnerability Details

File Location: prompts/reviewer.md:7-20
Vulnerability Type: Indirect prompt injection
Risk Level: Medium

Vulnerable Code

markdown
### 章节正文
{{chapter_content}}

### 风格配置
{{style_config}}

### 世界观索引
{{world_index}}

### 本章相关角色
{{characters}}

### 大纲信息
{{outline_info}}

Technical Analysis

The review prompt embeds chapter content, configuration, world data, character data, and outline information directly in the instruction context. It does not distinguish trusted reviewer instructions from untrusted project content or direct the model to disregard commands contained in that content.

A malicious chapter or project file can therefore include prompt-injection text that suppresses findings, changes issue counts, fabricates a favorable assessment, or replaces the required review structure.

Attack Path

  1. An attacker creates or modifies a chapter, character card, outline, world index, or style configuration.
  2. The modified content contains instructions targeting the reviewer, such as directing it to report zero problems.
  3. The review workflow interpolates the content into one of the template variables.
  4. The model follows the embedded instruction instead of, or in addition to, the trusted four-dimensional review procedure.
  5. The resulting report omits genuine issues or contains attacker-selected conclusions.
  6. Downstream users or automated workflows rely on the compromised review when approving or modifying the chapter.

Impact Assessment

Successful exploitation undermines the integrity and reliability of editorial review results. It can conceal consistency, logic, style, or scale-related issues and influence downstream project edits. The reviewed prompt does not itself provide system command execution, network access, sensitive-file access, or privilege escalation, so the demonstrated scope remains confined to review-output and related workfl ...[truncated 16 chars]

Remediation
View remediation

Remediation Suggestions

  • Add an explicit instruction that all interpolated material is evidence to review, not a source of executable instructions.
  • Wrap each input in unique data delimiters and require the reviewer to ignore embedded attempts to change its role, criteria, or output format.
  • Validate structured project data against strict schemas before interpolation.
  • Keep trusted review criteria in a higher-priority message than project-controlled content when supported by the runtime.
  • Detect and flag instruction-like text in reviewed content without treating detection alone as a complete defense.
  • Require human verification before automatically accepting a review or applying its recommendations.
Vulnerability Patterns
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • 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 (15)

Hidden Instructions

High
Category
Prompt Injection
Confidence
60% confidence
Finding

Hidden instructions were detected in comments or invisible text. These could contain malicious directives. Manual review is recommended.

Content

Scanner excerpt · SKILL.md (reported line 1)May include surrounding context.

md
# 风格分析 Prompt

你是一位专业的文学风格分析师。请分析以下参考文章,提取风格特征。

Hidden Instructions

High
Category
Prompt Injection
Confidence
60% confidence
Finding

Hidden instructions were detected in comments or invisible text. These could contain malicious directives. Manual review is recommended.

Content

Scanner excerpt · prompts/deviation-checker.md (reported line 1)May include surrounding context.

md
# 风格分析 Prompt

你是一位专业的文学风格分析师。请分析以下参考文章,提取风格特征。

Hidden Instructions

High
Category
Prompt Injection
Confidence
60% confidence
Finding

Hidden instructions were detected in comments or invisible text. These could contain malicious directives. Manual review is recommended.

Content

Scanner excerpt · prompts/outline-generator.md (reported line 1)May include surrounding context.

md
# 风格分析 Prompt

你是一位专业的文学风格分析师。请分析以下参考文章,提取风格特征。

Hidden Instructions

High
Category
Prompt Injection
Confidence
60% confidence
Finding

Hidden instructions were detected in comments or invisible text. These could contain malicious directives. Manual review is recommended.

Content

Scanner excerpt · prompts/reviewer.md (reported line 1)May include surrounding context.

md
# 风格分析 Prompt

你是一位专业的文学风格分析师。请分析以下参考文章,提取风格特征。

Hidden Instructions

High
Category
Prompt Injection
Confidence
60% confidence
Finding

Hidden instructions were detected in comments or invisible text. These could contain malicious directives. Manual review is recommended.

Content

Scanner excerpt · prompts/segment-writer.md (reported line 1)May include surrounding context.

md
# 风格分析 Prompt

你是一位专业的文学风格分析师。请分析以下参考文章,提取风格特征。

Hidden Instructions

High
Category
Prompt Injection
Confidence
60% confidence
Finding

Hidden instructions were detected in comments or invisible text. These could contain malicious directives. Manual review is recommended.

Content

Scanner excerpt · prompts/style-analyzer.md (reported line 1)May include surrounding context.

md
# 风格分析 Prompt

你是一位专业的文学风格分析师。请分析以下参考文章,提取风格特征。

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

Generic trigger phrases such as '写小说', '创建小说', '生成大纲', and '审读' are common enough to overlap with ordinary conversation and other writing tasks. This can lead to accidental invocation, causing the agent to apply file operations or specialized prompts in contexts the user did not intend.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

The trigger condition '任何与小说创作相关的请求' is an overly broad catch-all that can cause the skill to activate for vague or incidental mentions of novels. Over-broad activation increases the chance of unintended context capture or workflow takeover when the user did not explicitly request this skill.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

This markdown prompt forces a specific language/locale for the skill's interaction by presenting all instructions and output format in Chinese only. The file does not provide any user opt-in, alternative language option, or documented region-specific justification, which fits the natural-language policy violation criteria.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The prompt is written entirely in Chinese and instructs the model to produce the outline in that language context, but it does not offer any language or locale choice to the user. Under the policy, forcing a specific language without user opt-in is a natural-language policy violation unless the constraint is explicitly justified.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The prompt explicitly instructs the model to '用简体中文写作', which imposes a specific language/locale requirement. Under the policy, language constraints are only acceptable when the skill offers user choice or clearly documents a justified region-specific need, 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
97% confidence
Finding

The prompt is entirely written in Chinese and instructs the model to analyze and output in that language, but it does not provide any user opt-in or alternative locale choice. Under the language/locale policy, forcing a specific language without user selection is a natural-language policy violation unless clearly justified as region-specific.

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 is entirely in Chinese and provides no language selection, translation, or documented locale restriction. This can prevent reviewers, users, and safety systems from understanding the skill content, which is especially risky here because the template contains domination/dehumanization themes that may evade oversight when not broadly accessible.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
95% confidence
Finding

The note "默认简体中文" imposes a language default in natural-language instructions, but the file does not explicitly offer users a language selection or state that the language can be changed on request. Under the policy, forcing a specific language without user opt-in can be a locale/language policy issue.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
95% confidence
Finding

The prompt title and all instructions are written in Chinese and direct the model to act as a Chinese-language reviewer, with no indication that another language may be used on user request. This is a natural-language locale constraint and may violate language-choice policy when no opt-in or justification is provided.

Content

No source excerpt is available for this finding.