Back to skill

Security audit

circuit-intimate-therapy

Security checks for vulnerabilities and agentic risk

Overview

The skill is mostly coherent, but it requires saving sensitive relationship-case details into local HTML files without clear consent, minimization, or retention controls.

Review before installing. This skill does not show exfiltration or malicious code, but it is designed to create local HTML files containing details from sensitive relationship discussions. Install only if users understand where files are written, can decline generation, and are comfortable managing or deleting those files themselves.

Vulnerability Patterns
  • 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
  • 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
Findings (1)

T09 · Insecure Skill Coding Practices

Warning
Location
.html`(浏览器双击打开)。简短一句话说这个 app 演示什么、关键控件是什么。 ``` From `references/app-generation-template.md:290-309`: ```markdown ## 上下文嵌入清单(让 app 真的"和本次对话相关") **生成的 app 不该是泛化模板,要在 5 处嵌入本次对话的具体上下文:** | 嵌入点 | 怎么做 | |---|---| | 标题 + subtitle | 写明本次讨论的具体问题 / 案例 / 来访者代号 | | 默认参数值 | 用本次诊断推出的 R/L/C/DC 起始值,而非 canonical default | | 预设按钮 | 加"当前态" / "目标态" / "X 个月后预测",值来自对话 | | 诊断文字(.insight) | 引用对话里的具体语句、依恋类型、Gottman 类别 | | 案例标签 | 如果是案例咨询,在某处显眼放"为 [来访者代号 / 主题] 定制" | **反例**(不要这样做): - 直接复制 app01 改个标题 - 滑块默认值 0.4/1.0/1.0(完全是 app01 ...[truncated 3570 chars]:114
Finding

Mandatory Persistence of Sensitive Conversation Data Outside the Skill Workspace

Content
View full analysis
.html ``` - 必须在 `circapps/generated/` 子目录(与 12 个 canonical app 物理隔离) - 文件名 kebab-case 反映本次核心概念,不要含日期(例 `volatile-to-validator.html`、`zener-threshold-tuner.html`) - 同名直接覆盖(连续追问同主题就在迭代同一个 app) ### 生成流程(每次) 1. **判断该不该生成**(见下文跳过条件) 2. **选原型**(详见 `references/app-generation-template.md`): - 时域响应 → 借鉴 app01 / app06 - 频率响应 → 借鉴 app02 / app05 - Z 平面定位 → 借鉴 app03 - I-V 曲线 → 借鉴 app07 - 多面板对比 → 借鉴 app04 / app09 3. **嵌入本次对话上下文**(必须 ≥ 3 处): - 标题 / subtitle 写明本次具体讨论的问题或案例 - 默认参数值 = 本次诊断推出的 R/L/C/DC,而非泛化 default - 预设按钮 = "当前态" / "目标态" / 案例特定预设 - `.insight` 诊断文字引用对话里的具体语句、依恋类型、Gottman 类别 4. **强制约束**: - ``(注意 `../`) - ``(注意 `../`) - **不要调** `setNav()` —— 它写死 `href="index.html"` 在子目录会 404;手写一个 `← 索引` 的 nav - 单页 HTML 自洽,不引入除 common.css/js 外的依赖 - canvas 用 `setupC ...[truncated 4087 chars]
Remediation
View remediation
Vulnerability Patterns
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • System Prompt LeakageDirect Leakage, Indirect Extraction, Tool-Based Exfiltration
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
Findings (23)

Hidden Instructions

High
Category
Prompt Injection
Confidence
70% confidence
Finding

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

Content

Scanner excerpt · references/app-generation-template.md (reported line 58)May include surrounding context.

md
<h1><完整标题></h1>
<p class="subtitle"><一两句:这个 app 演示什么、和本次对话哪个概念对应></p>

<!-- 控件区 + canvas 区,用 .layout-2col 或 .layout-3col -->

</div>

Direct Prompt Extraction

High
Category
System Prompt Leakage
Confidence
85% confidence
Finding

Skill contains instructions that could directly expose system prompts, internal rules, or hidden instructions to users or external parties.

Content

Scanner excerpt · references/apps-guide.md (reported line 181)May include surrounding context.

md
|---|---|
| **Social penetration theory 社会渗透** | 自我表露的导通阈值:轻度披露不触发互惠,只有跨过亲密阈值才开启互惠流动 |
| **Triggering events** | 冲突点火阈值:日常小摩擦在阈值下被吸收 |
| **Microexpressions 微表情** | **次阈值漏电流**:在 display rules 关门前的瞬间漏过去 |
| **Hurt feelings 感情创伤** | **非对称阈值**:负向输入的导通阈值远低于正向(损失厌恶在亲密关系里的版本) |

### 临床用途

Direct Prompt Extraction

High
Category
System Prompt Leakage
Confidence
85% confidence
Finding

Skill contains instructions that could directly expose system prompts, internal rules, or hidden instructions to users or external parties.

Content

Scanner excerpt · references/diodes.md (reported line 45)May include surrounding context.

md
|---|---|
| **Social penetration theory 社会渗透** | 自我表露的导通阈值:轻度披露不触发互惠,只有跨过亲密阈值才开启互惠流动 |
| **Triggering events** | 冲突点火阈值:日常小摩擦在阈值下被吸收 |
| **Microexpressions 微表情** | **次阈值漏电流**:在 display rules 关门前的瞬间漏过去 |
| **Hurt feelings 感情创伤** | **非对称阈值**:负向输入的导通阈值远低于正向(损失厌恶在亲密关系里的版本) |

### 临床用途

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
98% confidence
Finding

Forcing all responses to Chinese without user opt-in can undermine user awareness and informed consent, especially around side effects, safety warnings, or file-writing behavior. This is not a direct code-execution issue, but it can increase the chance that a user misses important disclosures or misunderstands what the agent is doing.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The skill requires the agent to write a new HTML file to disk on every substantive response, creating a persistent side effect that the user may not have clearly requested or understood. Even though the target directory is constrained, silent file creation can fill storage, overwrite prior generated content, and normalize tool use without explicit consent, which is risky in an agent setting.

Content

No source excerpt is available for this finding.

Ae4

Medium
Category
analysis-evasion
Confidence
80% confidence
Finding

Suspicious Unicode normalization or mixed-script content

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The skill requires the agent to use a Write tool to create files on the local filesystem whenever it answers, but does not require explicit user confirmation before modifying local files. This creates an unsafe side effect boundary: a user asking for advice may unknowingly trigger file creation on disk.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

The template explicitly instructs overwriting same-named files without warning or confirmation. This can destroy prior user work or silently replace a previously reviewed artifact with a new one, especially because the agent is encouraged to iterate on reused filenames.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
98% confidence
Finding

The template hard-codes <html lang="zh">, which imposes a specific language/locale choice on all generated apps. The document does not indicate that this should vary based on user preference or context, so it constitutes a language policy issue.

Content

No source excerpt is available for this finding.

Ssd 3

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The skill instructs the agent to embed concrete conversation details, client identifiers, and case-specific facts into generated HTML files stored on disk. Because this skill is for intimate-relationship counseling, those details are highly sensitive; persisting them locally increases the risk of privacy leakage, accidental disclosure, and long-term retention of sensitive personal data.

Content

No source excerpt is available for this finding.

Ae4

Medium
Category
analysis-evasion
Confidence
80% confidence
Finding

Suspicious Unicode normalization or mixed-script content

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The document title and all operational guidance are written entirely in Chinese, which constitutes a language constraint in the skill's natural-language instructions. There is no indication that users may choose another language or that the locale restriction is region-specific and justified.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
89% confidence
Finding

The file hard-codes China-specific crisis resources and a Chinese-language locale path without any user-location or language check. In a safety-sensitive counseling workflow, this can misdirect users to irrelevant or inaccessible resources, delay access to appropriate help, and exclude users who do not read Chinese.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The entire skill file, including headings, usage instructions, and examples, is written only in Chinese and instructs behavior for responding to users in that language. There is no indication that users may choose another language or that the Chinese-only constraint is required for a region-specific or compliance reason.

Content

No source excerpt is available for this finding.

Ae4

Medium
Category
analysis-evasion
Confidence
80% confidence
Finding

Suspicious Unicode normalization or mixed-script content

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The trigger guidance uses broad, everyday relationship questions as activation cues without clear scope limits, exclusions, or safety gating. In a counseling-oriented skill that also generates artifacts and steers conceptual framing, overbroad triggers can cause unintended activation, irrelevant therapeutic framing, and use in contexts where the analogy is inappropriate or potentially harmful.

Content

No source excerpt is available for this finding.

Ae4

Medium
Category
analysis-evasion
Confidence
80% confidence
Finding

Suspicious Unicode normalization or mixed-script content

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
76% confidence
Finding

The document is written entirely in Chinese and does not state that language selection is optional or limited to a specific region or audience. Under the policy, forcing a specific language without user opt-in can be a natural-language policy violation unless the locale constraint is explicitly justified.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

This file gives detailed, prescriptive relationship and family intervention guidance in a quasi-clinical framing, including prioritization and decision rules, before clearly foregrounding that it is not a substitute for professional or crisis care. In a mental-health-adjacent skill, users may over-rely on the model for real-world treatment decisions, delaying appropriate licensed, emergency, or culturally competent support.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The document presents all user-facing instructional content in Chinese and does not indicate that the user can select another language. Under the policy, forcing a specific language without user opt-in 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

Low
Category
Not specified by scanner
Confidence
84% confidence
Finding

The title and core instructions are written to require a fixed Chinese-oriented generation pattern, including Chinese phrasing in the final output example. There is no statement that the language should adapt to the user's preference, which can conflict with organizational language-choice policies.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
85% confidence
Finding

All instructional content and trigger phrases are presented exclusively in Chinese, and the file does not indicate that the skill supports user language preference or is intentionally limited to a Chinese-speaking audience. Under the policy, forcing a specific language without opt-in can be a natural-language policy issue unless the locale limitation is documented and justified.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
89% confidence
Finding

This markdown file contains user-facing natural-language instructions and content exclusively in Chinese, with no indication that the language choice is optional or region-specific. Under the policy rule, forcing a specific language without user opt-in is a natural-language policy concern.

Content

No source excerpt is available for this finding.