Back to skill

Security audit

fusion-search

Security checks for vulnerabilities and agentic risk

Overview

This is a real search skill, but it uses stealth browsing and unsafe Chromium settings while fetching arbitrary result pages without enough user control or network safeguards.

Install only in a contained environment with restricted outbound network access, especially blocking private networks and metadata endpoints. Avoid sending secrets or internal queries, and treat stealth browsing plus full-content fetching as behavior that may have policy or site-terms implications.

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 (3)

T09 · Insecure Skill Coding Practices

Error
Location
scripts/fusion_search.py:157
Finding

Unrestricted Server-Side Navigation Enables SSRF

Content
View full analysis
{ for (const s of document.querySelectorAll( 'script, style, nav, header, footer, .ad, .sidebar, ' + '.comment, .popup, .modal, .cookie, .advertisement, ' + 'noscript, iframe, .related, .recommend, .share' )) s.remove(); }""") # 优先找主要内容区域 content = page.evaluate("""() => { const main = document.querySelector('article, main, .content, ' + '.post, .article, #content, #main, .entry-content, ' + '.post-content, [role="main"]'); if (main) return main.innerText; return document.body ? document.body.innerText : ''; }""") content = re.sub(r'\s+', ' ', content).strip() return content[:8000] except Exception: return "" ``` The URL is taken directly from search results: ```python for i, r in enumerate(results[:count]): content = fetch_full_content(page, r["url"], timeout=10000) if content: results[i]["content"] = content ``` ### Technical Analysis The full-content extraction feature navigates to search-result URLs without validating: - The URL scheme - The destination hostname - The resolved IP address - Redirect destinations - Loopback, private, link-local, or reserved address ranges - Cloud metadata endpoints - DNS ...[truncated 2264 chars]
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Error
Location
scripts/fusion_search.py:39
Finding

Chromium Sandbox and Web Security Protections Are Disabled

Content
View full analysis
Remediation
View remediation

T08 · Insecure Dependencies

Note
Location
metadata.json:15
Finding

Playwright and Chromium Dependencies Are Not Version-Pinned

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
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • MCP Least PrivilegeUnderdeclared Capability, Wildcard Permission, Missing Permission Declaration
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
Findings (12)

Tp4

High
Category
MCP Tool Poisoning
Confidence
90% confidence
Finding

The documented behavior substantially overstates what the implementation appears to do, including anti-bot browsing, multi-engine routing, and full-content fetching. This mismatch is dangerous because reviewers and users may grant trust, permissions, or operational approval based on inaccurate claims, obscuring the real attack surface and preventing effective risk assessment.

Content

No source excerpt is available for this finding.

Vague Triggers

High
Category
Not specified by scanner
Confidence
97% confidence
Finding

The trigger keywords include very generic terms such as '搜索', '最新', '是什么', 'search', and 'find', which are likely to activate during ordinary conversation. Over-broad triggering can cause unintended network searches or content fetching, exposing user prompts to external services and invoking the skill in contexts where the user did not explicitly consent.

Content

No source excerpt is available for this finding.

Undeclared Tool Scope

Medium
Category
MCP Least Privilege
Confidence
92% confidence
Finding

The skill declares a network-capable search tool and references file-backed scripts, but the manifest does not constrain permissions or allowed tools. In practice this weakens least-privilege controls, making it easier for the skill to perform broader network or local file interactions than a reviewer or runtime policy might expect.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The skill advertises external search and optional full-page content retrieval but does not clearly warn that user queries and fetched URLs will be transmitted to third-party services. This creates privacy and compliance risk because sensitive prompts, internal terms, or browsing targets could be disclosed outside the local environment without informed consent.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
90% confidence
Finding

The skill describes stealth.js-based anti-detection browsing and anti-crawling evasion without clearly warning operators about the resulting network and policy implications. Anti-detection behavior can bypass normal expectations of transparent automation, potentially violating site terms, triggering security monitoring issues, or creating legal and reputational exposure for the deploying environment.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The Bing CN URL forces mkt=zh-CN, setlang=zh-CN, and cc=CN, and the file overall defines multiple engines with fixed language values such as zh and en. This is a natural-language locale policy concern because the skill imposes language/region behavior rather than offering a user choice or documenting an explicit region-specific limitation.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The Playwright browser context is hard-coded to use locale="zh-CN" and timezone_id="Asia/Shanghai", which imposes a specific language/locale behavior on all searches. This matches the policy category for language or locale constraints because the file provides no user opt-in or alternative locale selection.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The secondary browser context used for fetching full page content again hard-codes locale="zh-CN". Repeating the locale restriction in this path means the entire skill consistently enforces a specific locale rather than letting the user choose.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The code classifies any query containing Chinese characters as "zh" and all others as "en", creating a hard language/locale decision with no user choice. This can violate language/locale policy because the skill behavior is forced by heuristic detection rather than explicit user opt-in or a documented justified regional scope.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The URL builder defaults the language parameter to "zh-CN", which imposes a specific locale on generated search requests unless callers override it. This is a natural-language policy concern because it forces a locale preference without visible user opt-in in this file.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

This script deliberately alters multiple browser fingerprint surfaces (webdriver, plugins, languages, platform, hardware, permissions, canvas, and WebGL) to evade bot and anti-automation detection. In a search skill using Playwright and explicitly advertising anti-crawling bypass, this meaningfully increases the ability to access services in ways operators may not expect, creating abuse, compliance, and account/IP blocking risk.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

Forcing navigator.languages to a Chinese-first locale overrides the real client profile and can be used to influence content delivery while also contributing to fingerprint spoofing. In this context, it makes the stealth behavior more targeted and deceptive, potentially causing region-specific scraping behavior, misleading downstream services, and inconsistent or biased results without user awareness.

Content

No source excerpt is available for this finding.