Back to skill

Security audit

Browser Automation Cdp

Security checks for vulnerabilities and agentic risk

Overview

The skill is mainly for browser automation, but it includes under-scoped control of logged-in browser sessions and unsafe guidance around cookies and browser processes.

Review before installing. Use this only with a dedicated browser profile for sites you explicitly want automated, avoid banking/admin/personal accounts, do not permit cookie/token/browser-profile extraction, confirm before screenshots or file exports, and close any remote-debugging browser when finished.

Vulnerability Patterns
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs
  • 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 (2)

T05 · Unauthorized Access and Privilege Escalation

Error
Location
SKILL.md:304
Finding

Direct Access to the Browser Cookie Database Bypasses Page-Level Security Boundaries

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, line 304
Vulnerability Type: Least-privilege violation and sensitive credential access
Risk Level: High

Complete documentation excerpt translated into English:

text
document.cookie cannot obtain HttpOnly cookies. If complete cookies are required, consider reading the SQLite database directly from the browser files (the file-locking issue must be handled).

Technical Analysis

The Skill recommends directly reading the browser cookie SQLite database when HttpOnly cookies cannot be accessed through document.cookie. HttpOnly is a deliberate browser security boundary intended to prevent page scripts from reading authentication cookies. Accessing the browser's underlying cookie database circumvents that page-level restriction and exposes credential material outside the scope of normal DOM automation.

Complete browser cookies can include authenticated session tokens for multiple unrelated services in the selected browser profile. Direct database access is not necessary for the Skill's declared navigation, clicking, screenshot, and DOM-extraction functions and therefore exceeds the minimum privileges required.

The instruction does not itself transmit cookies or provide a complete extraction implementation. Nevertheless, an Agent with the declared read and exec tools could use the recommendation to locate, copy, unlock, decrypt, and inspect browser credential storage.

Attack Path

  1. A user invokes the Skill for an authenticated browser-automation task.
  2. The target operation encounters an HttpOnly cookie that cannot be read through page JavaScript.
  3. Following the Skill's guidance, the Agent locates the browser profile's cookie SQLite database.
  4. The Agent copies or otherwise handles the locked database and attempts to read or decrypt its contents.
  5. Session cookies for the target site, and potentially unrelated sites in the same profi ...[truncated 701 chars]
Remediation
View remediation

Remediation Suggestions

  • Remove the recommendation to read browser cookie databases directly.
  • Explicitly prohibit extracting, decrypting, copying, displaying, or transmitting authentication cookies and browser credential stores.
  • Limit automation to page-scoped CDP operations required for navigation, interaction, screenshots, and approved DOM extraction.
  • Use a dedicated automation browser profile containing only the accounts needed for the specific task.
  • Require explicit, informed user approval before any credential-sensitive operation.
  • Apply domain allowlists and prevent access to cookies belonging to sites outside the active task.
  • Redact session identifiers and other credentials from logs, tool output, screenshots, and generated files.
  • If authenticated API access is genuinely required, use a narrowly scoped, revocable token supplied through an approved secret-management mechanism instead of browser session extraction.

T09 · Insecure Skill Coding Practices

Warning
Location
SKILL.md:80
Finding

Forceful Termination of All Edge Browser Processes Can Disrupt Unrelated User Activity

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, lines 80-84; related process-termination guidance at line 299
Vulnerability Type: Unsafe process termination and denial of service
Risk Level: Medium

Complete relevant command snippet:

powershell
taskkill /F /IM msedge.exe /T
Start-Sleep 3
Start-Process "C:\Program Files (x86)\Microsoft\Edge\Application\msedge.exe" --remote-debugging-port=9222

Start-Process "C:\Program Files (x86)\Microsoft\Edge\Application\msedge.exe" --remote-debugging-port=9222 --user-data-dir="C:\SkillPlatformBrowser\Edge"

Related troubleshooting command:

powershell
taskkill /F /PID <pid>

Technical Analysis

The startup procedure uses /F, /IM msedge.exe, and /T to forcefully terminate every Edge process and its child processes. It does not verify that a process belongs to the automation profile or that the user has saved active work. This is broader than necessary to start a dedicated CDP-enabled browser instance.

The troubleshooting guidance also recommends forcefully terminating a process by PID. Without validation of the executable path, owner, command line, and relationship to the conflicting port, an incorrect or stale PID could target an unrelated process.

The browser is then launched with a remote-debugging port. Binding details and access restrictions are not specified in the documented commands. The primary confirmed issue, however, is the indiscriminate forced termination; the audit did not verify that the debugging endpoint is remotely exposed.

Attack Path

  1. A user invokes the Skill while ordinary Edge windows contain active tabs, downloads, forms, or unsaved work.
  2. The Agent follows the documented setup instructions.
  3. taskkill /F /IM msedge.exe /T forcibly terminates all Edge processes and their descendants.
  4. Unrelated browser sessions are interrupted and unsaved state may be lost.
  5. During troubleshooting, an ...[truncated 506 chars]
Remediation
View remediation

Remediation Suggestions

  • Remove the blanket taskkill /F /IM msedge.exe /T instruction.
  • Launch automation with a dedicated browser profile and record the PID of the specific process created for the task.
  • Reuse an existing dedicated debugging instance when possible.
  • Request explicit user confirmation before terminating any browser or service process.
  • Attempt graceful shutdown before using forced termination.
  • Before terminating a PID, validate its executable path, owner, command line, start time, and actual ownership of the conflicting port.
  • Terminate only the dedicated automation process and avoid recursive child-process termination unless strictly required.
  • Warn users about unsaved work and active downloads before any shutdown operation.
  • Bind the CDP endpoint to loopback where supported and prevent untrusted local or network clients from reaching it.
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
Findings (7)

Missing User Warnings

High
Category
Not specified by scanner
Confidence
92% confidence
Finding

The skill is explicitly designed to control an already logged-in browser session, but the description does not prominently warn that actions and extracted data occur under the user's authenticated identity. In this context, lack of a clear warning materially increases the chance of unauthorized access to private account data, posting actions, or session misuse.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
88% confidence
Finding

Nearly all user-facing descriptive and instructional content in the skill file is presented only in Chinese, with no indication that language selection is optional or region-specific. This can violate language or locale policy where skills should not impose a single language without user opt-in or documented justification.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
88% confidence
Finding

The skill claims that risk code was removed and security enhanced, yet later documentation includes guidance that facilitates extraction of sensitive browser session material. This contradiction can create unjustified trust and lower reviewer or user caution around a skill that operates inside authenticated browser contexts.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The listed trigger keywords include very generic terms such as "chrome", "devtools", "执行", "automation", and "browser", which overlap with common discussion about browsers and automation rather than uniquely identifying this skill. The description does not provide clear activation constraints or negative examples, increasing the risk of unintended invocation.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
85% confidence
Finding

The screenshot example captures page contents and saves them to disk without any privacy or sensitivity warning. Because the skill operates in logged-in sessions, screenshots may include account details, messages, billing data, or other sensitive content that can persist locally or be exfiltrated later.

Content

No source excerpt is available for this finding.

Ssd 3

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The skill includes detailed instructions for extracting account/quota information from authenticated pages by reading full-page text and discusses escalating to cookie/database access when normal browser APIs are insufficient. In the context of CDP control over a logged-in browser, this materially enables sensitive data harvesting from private accounts beyond ordinary browsing automation.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The documentation explicitly suggests reading browser cookie SQLite files to obtain cookies that are intentionally inaccessible to normal web content via HttpOnly. In a skill whose core feature is operating through an already logged-in browser, this expands from browser automation into credential/session extraction, enabling account takeover or unauthorized access if followed.

Content

No source excerpt is available for this finding.