Back to skill

Security audit

Agent's Backyard Garden

Security checks for vulnerabilities and agentic risk

Overview

The skill is mostly a coherent garden website, but it sends visitor messages to an under-disclosed temporary external tunnel and has confirmed unsafe rendering paths for untrusted content.

Install only if you are comfortable with a social website skill that uses external services. Do not put sensitive personal information in the guestbook, prefer a pinned installer or verified source checkout, and treat the live site as needing security fixes before relying on it for public visitor content.

Vulnerability Patterns
  • 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
  • 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
Findings (6)

T09 · Insecure Skill Coding Practices

Error
Location
index.html:247
Finding

Stored Cross-Site Scripting Through Guestbook and GitHub Messages

Content
View full analysis
({ name: m.author || '匿名', message: m.content || '', time: m.time || '', source: 'web', })); ``` ```javascript // functions/api/messages.js return issues .filter(issue => !issue.pull_request && /留言|来访|hello|hi|你好/i.test(issue.title) ) .map(issue => ({ name: issue.user?.login || 'GitHub 访客', message: issue.title + (issue.body ? `\n${issue.body}` : ''), time: issue.created_at, source: 'github', url: issue.html_url, })); ``` ```javascript // index.html const res = await fetch('/api/messages'); if (!res.ok) throw new Error('not ready'); const data = await res.json(); if (data && data.messages && data.messages.length > 0) { wall.innerHTML = '
' + data.messages.map(m => '
' + '
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Error
Location
index.html:159
Finding

Remote Status Service Can Inject Executable HTML Into the Main Page

Content
View full analysis
({ name: p.name, text: p.status || '', })), }; return new Response(JSON.stringify(mapped), { headers: { 'Content-Type': 'application/json', ...CORS }, }); } return new Response(JSON.stringify(data), { headers: { 'Content-Type': 'application/json', ...CORS }, }); } } catch (e) { // Fall back to static data } } ``` ```javascript // index.html (async function loadStatus() { try { const res = await fetch('/api/status'); if (!res.ok) throw new Error('status not ready'); const data = await res.json(); if (data && data.statuses) { const board = document.getElementById('statusBoard'); board.innerHTML = data.statuses.map(s => '
' + s.name + '' + s.text + '
' ).join(''); } } catch (e) { console.log('Status API unavailable; using default data'); } })(); ``` ### Technical Analysis The status API trusts data obtained from an ephemeral `trycloudflare.com` tunnel. Both `p.name` and `p.statu ...[truncated 1536 chars]
Remediation
View remediation

other

Warning
Location
index.html:194
Finding

Guestbook Content and Visitor Metadata Are Sent to an Undisclosed Temporary Tunnel

Content
View full analysis
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Warning
Location
functions/api/guestbook.js:42
Finding

Failed Guestbook Persistence Is Falsely Reported as Successful Receipt

Content
View full analysis
document.getElementById('formSuccess').style.display = 'none', 4000); } else { document.getElementById('formSuccess').textContent = '留言已收到,稍后同步到留言墙 ✨'; document.getElementById('formSuccess').style.display = 'block'; document.getElementById('guestName').value = ''; document.getElementById('guestMessage').value = ''; setTimeout(() => document.getElementById('formSuccess').style.display = 'none', 4000); } ``` ### Technical Analysis When forwarding fails, the server does not save the message to a local database or durable queue. Nevertheless, it returns HTTP `201` with `success: true` and claims that the message was received for later synchronization. The browser also claims that a failed submission was received and clears the user's input. No later synchronization mechanism is present in the reviewed code. This violates message integrity and prevents users from distinguishing durable accept ...[truncated 767 chars]
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Warning
Location
scripts/build-articles.mjs:119
Finding

Markdown Builder Generates Unsanitized HTML and Unsafe URL Attributes

Content
View full analysis
$1'); // **bold** r = r.replace(/\*\*([^*]+)\*\*/g, '$1'); // *italic* r = r.replace(/\*([^*]+)\*/g, '$1'); // [text](url) r = r.replace(/\[([^\]]+)\]\(([^)]+)\)/g, '$1'); // Images r = r.replace(/!\[([^\]]*)\]\(([^)]+)\)/g, '$1'); return r; } ``` ```javascript const titleMatch = body.match(/^# (.+)$/m); const title = titleMatch ? titleMatch[1] : path.basename(mdFile, '.md'); const cleanTitle = inlineMdToHtml(title); const author = fm.owner || '后花园'; const created = fm.created || ''; const bodyWithoutH1 = body.replace(/^# .+\n*/, '').trim(); const htmlContent = mdToHtml(bodyWithoutH1); const result = TEMPLATE .replaceAll('${title}', cleanTitle) .replace('${author}', author) .replace('${series}', fm.domain || '后花园') .replace('${content}', htmlContent); ``` ### Technical Analysis `inlineMdToHtml()` performs regular-expression substitutions without first HTML-encoding ordinary text. As a result, raw HTML in Markdown can pass into generated pages unchanged. Link labels, image alternative text, URLs, titles, authors, and series values are also inserted without contextual output encoding. The builder does not allowlist URL protocols. Crafted Markdown can therefore generate dangerous or malformed `href` and `src` attributes. Frontmatter values can also inject markup into template locations, depending on the surrounding template context. The exploit requires an attacker to influence article Markdown or frontmatter and have the build script run. This can occu ...[truncated 975 chars]
Remediation
View remediation

T08 · Insecure Dependencies

Warning
Location
SKILL.md:71
Finding

Installation Instructions Invoke an Unpinned npm Tool

Content
View full analysis
Remediation
View remediation
Vulnerability Patterns
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
Findings (62)

Tp4

High
Category
MCP Tool Poisoning
Confidence
97% confidence
Finding

The declared description presents this skill as a place-like interactive agent community experience with six distinct agents and garden-related triggers. The supplied code does not implement that experience. Instead, it is narrowly a backend guestbook submission API that accepts POST requests and forwards messages to an external service. While a guestbook is mentioned in the description, the actual code only covers one supporting component of that larger concept and introduces an undeclared external data-forwarding behavior. Because the implemented primary function materially differs from the declared experiential purpose, this is a mismatch.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
95% confidence
Finding

The declared description emphasizes a narrative/interactive virtual community inhabited by six distinct agents, triggered by phrases like 'visit the garden.' However, the supplied code does not implement agent personas, stories, community interaction logic, or trigger handling. Its concrete function is a backend API endpoint for retrieving and merging visitor messages from a remote API and GitHub Issues. While guestbook/messages are loosely related to a 'community,' the primary behavior is materially narrower and different from the declared purpose. The code also accesses external resources (a trycloudflare endpoint and GitHub API) that are not reflected in the declaration.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
97% confidence
Finding

The declared description presents the skill as a narrative/interactive destination where users can 'visit the garden' and encounter six agents, stories, and a guestbook. The supplied code does not implement that experience. Instead, it is a backend status endpoint that returns JSON agent statuses, optionally pulling them from an external API and remapping fields. While the six agent names overlap with the description, the primary purpose is materially different: status retrieval/service integration rather than a garden/community experience. The external fetch and API-serving behavior are undeclared capabilities, and the described triggers and guestbook/story functionality are absent from this code chunk.

Content

No source excerpt is available for this finding.

Lp1

High
Category
MCP Least Privilege
Confidence
93% confidence
Finding

This endpoint performs outbound network access to a third-party service, but the capability is not declared or justified in the skill metadata. Undeclared network egress is dangerous because it enables silent data transmission outside the expected trust boundary and weakens permission-based review of what user data leaves the skill.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

High
Category
Not specified by scanner
Confidence
98% confidence
Finding

The form posts user-supplied content directly to a hard-coded external Cloudflare tunnel endpoint, which is an unstable third-party exposure point and bypasses same-origin expectations. In a skill framed as a passive 'garden' experience, this expands data exfiltration risk because user text is transmitted to infrastructure not clearly disclosed or justified.

Content

No source excerpt is available for this finding.

Rp1

Medium
Category
MCP Rug Pull
Confidence
90% confidence
Finding

The README instructs users to install the skill via npx skills add YuanZhiTech/agent-garden without pinning a specific version or commit. That allows whatever package or remote artifact is current at install time to be fetched, creating a supply-chain risk if the upstream package changes, is compromised, or is replaced with malicious content.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The README instructs users to activate the skill by saying "去花园看看。" without clarifying scope, context, or exclusion conditions. This natural-language trigger is broad and conversational, which increases the chance of unintended invocation or overlap with ordinary speech.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
88% confidence
Finding

The usage instruction tells the user to speak a Chinese phrase to the agent and does not offer any language choice or note that other languages are supported. This can violate language/locale policy expectations when a skill imposes a specific language without explicit opt-in or justification.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The description lists explicit trigger phrases, but also adds 'or wants to experience an Agent community,' which is open-ended and subjective. That broad condition lacks clear boundaries for when the skill should or should not activate, increasing the chance of unintended invocation.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

The manifest frames the skill as a passive social/community destination, yet requests Write and Bash(curl:*) in addition to network-capable read behavior. This mismatch is dangerous because an agent or platform may grant unnecessary privileges, enabling file modification or arbitrary outbound requests under a benign-seeming trigger. The social framing makes over-privileging more concerning because users are less likely to expect side effects.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

Granting shell-based curl:* access is broader than necessary for a skill whose documented role is essentially to point users to a website and describe community content. Broad shell/network capability can be abused to contact arbitrary remote services, exfiltrate context, or retrieve untrusted payloads, especially if the skill later processes external content. The low-stakes social context makes such broad capability harder to justify.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

Granting shell-based curl:* access is broader than necessary for a skill whose documented role is essentially to point users to a website and describe community content. Broad shell/network capability can be abused to contact arbitrary remote services, exfiltrate context, or retrieve untrusted payloads, especially if the skill later processes external content. The low-stakes social context makes such broad capability harder to justify.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The phrase 'or similar' leaves the activation criteria undefined, so agents may infer overly broad matches from ordinary conversation. This is a classic vague-trigger pattern because it does not specify what counts as sufficiently similar or provide exclusions.

Content

No source excerpt is available for this finding.

Rp1

Medium
Category
MCP Rug Pull
Confidence
96% confidence
Finding

The article instructs users to install and execute an unpinned package via npx skills add YuanZhiTech/agent-garden. Because no version is pinned, future upstream changes or a compromised dependency/repository could cause different code to run than the reviewer or user expects, creating a supply-chain execution risk. In this skill context, that is more dangerous than a generic documentation snippet because the page explicitly encourages installation into an agent runtime.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The markdown instructs users to trigger the skill by saying “去花园看看。”, but it does not define when this phrase should or should not activate the skill, nor provide any negative examples or context limits. As written, the phrase is conversational and broad enough to risk unintended invocation in ordinary dialogue.

Content

No source excerpt is available for this finding.

Rp1

Medium
Category
MCP Rug Pull
Confidence
90% confidence
Finding

The guide instructs users to run npx skills add YuanZhiTech/agent-garden without pinning a version or commit, which means future installs may pull whatever code the publisher serves at that time. This creates a supply-chain risk: if the package, dependency chain, or referenced repository is compromised, users could install or execute unintended code or prompts.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The file presents an agent framed as a garden/community character, but extensively documents investment-assistant behavior, stock analysis workflow, memory practices, and trading-oriented decision support. This creates capability drift: a user entering a benign social/community experience may be steered toward financial guidance or operational trading behaviors that are outside the declared purpose and therefore outside expected trust boundaries and safety review.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

This section explicitly advertises financial-analysis and simulated-trading capabilities, including historical K-line retrieval, real-time quotes, and simulated trading management, despite the skill being described as a community/storytelling garden. In context, that mismatch is dangerous because it can normalize finance-related actions in an environment where users and downstream systems would not expect higher-risk financial functionality or corresponding safeguards.

Content

No source excerpt is available for this finding.

External Transmission

Medium
Category
Data Exfiltration
Confidence
50% confidence
Finding

Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.

Content

Scanner excerpt · articles/18-千里相会.html (reported line 137)May include surrounding context.

html
只需要改几个环境变量:

```bash
ANTHROPIC_BASE_URL=https://api.deepseek.com/anthropic
ANTHROPIC_MODEL=deepseek-v4-flash[1m]
```

External Transmission

Medium
Category
Data Exfiltration
Confidence
50% confidence
Finding

Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.

Content

Scanner excerpt · articles/18-千里相会.md (reported line 51)May include surrounding context.

只需要改几个环境变量:

bash
ANTHROPIC_BASE_URL=https://api.deepseek.com/anthropic
ANTHROPIC_MODEL=deepseek-v4-flash[1m]

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The code hard-codes a Cloudflare tunnel URL as the destination for guestbook submissions. A temporary tunnel endpoint is risky because it may be ephemeral, harder to govern, and outside normal production controls, increasing the chance of data leakage, service hijack, or forwarding to infrastructure that has not been security-reviewed.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

This code forwards guestbook content to an external API but provides no user-facing notice in the request path that names and messages will be transmitted off-service. Lack of transparency around data transmission undermines informed consent and can expose users to privacy harms if they share personal information in the guestbook.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
89% confidence
Finding

The guestbook API sends user-submitted names and messages to an external persistence service, which is more than a purely local 'community/place' experience implies. Because the transmitted content is user-generated and potentially sensitive, this creates a data-sharing risk if users are not expecting off-platform storage.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
88% confidence
Finding

The manifest presents the skill as a place-like community experience with stories and a guestbook, but this endpoint actually depends on two external data sources: a temporary Cloudflare-tunneled API and the GitHub Issues API. That makes the behavior materially broader than a self-contained community space, because it functions as a cross-service content aggregator.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The API sets Access-Control-Allow-Origin to '*' and returns aggregated visitor messages from external services, allowing any website to read and republish that data directly from a browser. While the data may be public, permissive CORS broadens unauthorized reuse, scraping, and correlation of community activity beyond the garden's intended context.

Content

No source excerpt is available for this finding.