Back to skill

Security audit

chrome-qa-agent

Security checks for vulnerabilities and agentic risk

Overview

ClawScan marked this skill suspicious because its browser testing guidance includes potentially consequential actions and broad permission recommendations, while parts of the artifact were only partially analyzed.

Consider the suspicious verdict a reason to review carefully before installing. The reported concerns include live or destructive browser actions, broad browser-tool preapproval, and remote scripts running in authenticated pages. Analysis was partial: 10 of 12 components were fully inspected and two were partially inspected, so the evidence does not establish that unreviewed content is safe.

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
Vulnerability Patterns
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Output HandlingUnvalidated Output Injection, Cross-Context Output, Unbounded Output
  • Memory PoisoningPersistent Context Injection, Context Window Stuffing, Memory Manipulation
Findings (18)

Account-deletion test confirms a destructive action without a disposable-data warning.

High
Category
Output Handling
Confidence
98% confidence
Finding

The example explicitly confirms account deletion without first requiring a disposable account or authorization for destructive testing. Although it uses localhost and an isolated browser context, neither guarantees that the backend or account data is disposable.

Content

No source excerpt is available for this finding.

Agent Config Directory Access

High
Category
Agent Snooping
Confidence
90% confidence
Finding

This is not merely a reference to a configuration path: the surrounding instructions recommend changing the user-wide settings file to persistently preapprove the browser server. That can extend weakened approval controls beyond the current QA task; it is the same underlying permission issue as the adjacent preapproval finding, not evidence of credential theft.

Content

Scanner excerpt · SKILL.md (reported line 185)May include surrounding context.

避免权限疲劳

本 skill 会发起大量 MCP 工具调用(快照、点击、eval)。为免逐个批准,在 .claude/settings.json(项目级)或 ~/.claude/settings.json(用户级)里放行 chrome-devtools 工具:

json
{

Adversarial form submissions lack a warning about modifying live data through authenticated sessions.

High
Category
Output Handling
Confidence
97% confidence
Finding

The execution workflow submits forms and tests duplicate submissions while the skill also permits deployed targets and inherited login sessions. It lacks a prerequisite that distinguishes disposable test operations from real transactions, so general QA authorization can become unintended authorization to modify live data.

Content

No source excerpt is available for this finding.

Incomplete referenced artifact analysis

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 493)May include surrounding context.

md
确切的 `evaluate_script` 配方见 [references/browser-recipes.md](references/browser-recipes.md)。

Incomplete referenced artifact analysis

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 546)May include surrounding context.

md
- **浏览器配方** —— [references/browser-recipes.md](references/browser-recipes.md) —— 跑确定性检查(axe-core、控制台、图片、表单标签)时加载

Incomplete referenced artifact analysis

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 554)May include surrounding context.

md
带确切命令的实战示例见 [EXAMPLES.md](EXAMPLES.md),如果你想看断言协议的实际运行效果。

Agent Config Directory Access

High
Category
Agent Snooping
Confidence
85% confidence
Finding

Skill reads from agent configuration directories (.claude/, .codex/, .gemini/). These directories may contain API keys, personal settings, and other credentials that the skill has no legitimate need to access.

Content

Scanner excerpt · references/design-consistency.md (reported line 12)May include surrounding context.

bash
# Read the design system
cat references/design-system.md
# or: cat .claude/skills/chrome-qa-agent/references/design-system.md

# Then screenshot the changed page and check against documented patterns
mcp__chrome-devtools__take_screenshot { filePath: ".context/ui-test-screenshots/changed-page.png" }

Anti-Refusal Statement

High
Category
Anti-Refusal
Confidence
80% confidence
Finding

Skill instructs the agent to omit warnings, disclaimers, or ethical commentary. Stripping safety caveats hides risk from the user and is a common jailbreak preamble.

Content

Scanner excerpt · references/exploratory-testing.md (reported line 78)May include surrounding context.

md
EVIDENCE: Clicked "Regenerate" and the key changed immediately with no warning

Accessibility audit loads third-party executable code into an authenticated page without a privacy warning.

Medium
Category
Output Handling
Confidence
97% confidence
Finding

The audit injects CDN-hosted JavaScript directly into an authenticated page without integrity verification. If that dependency or its delivery is compromised, the script can read sensitive page content and perform same-origin actions using the active session; version pinning alone does not prevent this.

Content

No source excerpt is available for this finding.

Context Window Stuffing

Medium
Category
Memory Poisoning
Confidence
80% confidence
Finding

Skill attempts to fill the context window with filler content, displacing legitimate instructions and safety constraints. This can degrade agent performance or bypass safety boundaries.

Content

Scanner excerpt · EXAMPLES.md (reported line 471)May include surrounding context.

md
# uid=4 textbox "Message", uid=2 textbox "Email", uid=9 button "Send"

# Extremely long input
mcp__chrome-devtools__fill { uid: "4", value: "This is a very long message that keeps going and going and going and going and going and going and going and going and going and going and going and going and going and going and going and going and going and going and going and going and going and going and going and going and going" }
mcp__chrome-devtools__take_snapshot
mcp__chrome-devtools__take_screenshot { filePath: ".context/ui-test-screenshots/long-input.png" }
# Check: does textarea grow? Overflow? Break layout?

Unbounded Output

Medium
Category
Output Handling
Confidence
60% confidence
Finding

Output size or generation rate is not bounded. Unbounded output enables denial-of-service through resource exhaustion, log flooding, or context-window stuffing.

Content

Scanner excerpt · EXAMPLES.md (reported line 496)May include surrounding context.

探索式发现的报告格式:

text
FINDING: Contact form textarea has no maxlength — 500+ char input accepted without truncation
SEVERITY: low
ROUTE: /contact
EVIDENCE: Filled 280 chars into message field, form accepted it, layout intact

The adversarial testing workflow lacks a warning about modifying real data on deployed sites.

Medium
Category
Output Handling
Confidence
92% confidence
Finding

The README combines adversarial submissions and rapid clicks with support for deployed sites using existing authenticated sessions, without defining a safe mutation boundary. Following this workflow can create duplicate transactions, alter real records, or submit test payloads to live systems.

Content

No source excerpt is available for this finding.

Broad browser-tool preapproval is recommended without warning about the security implications.

Medium
Category
Output Handling
Confidence
96% confidence
Finding

The skill recommends blanket preapproval of browser tools, including arbitrary page JavaScript and state-changing interactions, to avoid confirmations. With access to the user's authenticated Chrome, this removes an approval boundary that could otherwise interrupt harmful actions or prompt-injection-driven misuse.

Content

No source excerpt is available for this finding.

Dialog testing instructs the agent to confirm actions without warning about potentially destructive side effects.

Medium
Category
Output Handling
Confidence
95% confidence
Finding

The reusable pattern tells the agent to confirm a dialog and verify that its side effect occurred, without first identifying or authorizing that effect. Applied indiscriminately to interactive elements, it can approve deletion, credential rotation, or other irreversible operations.

Content

No source excerpt is available for this finding.

The accessibility recipe executes third-party CDN code in the inspected page without warning about its access to page data.

Medium
Category
Output Handling
Confidence
96% confidence
Finding

The recipe loads CDN-hosted JavaScript into the inspected page without integrity verification, granting that script access to the page's DOM, origin-accessible storage, and authenticated same-origin operations. Although axe-core is a legitimate testing library and its version is pinned, a compromised CDN asset could abuse those privileges; support for arbitrary deployed URLs in a local Chrome session increases the potential exposure.

Content

No source excerpt is available for this finding.

The interaction recipe includes form submissions without warning that they can change real application data.

Medium
Category
Output Handling
Confidence
90% confidence
Finding

The interaction recipe explicitly includes form submissions without distinguishing harmless test interactions from operations that modify real application state. The placeholder click is not itself destructive, but following this guidance against a deployed application in an authenticated browser could create records, send messages, or trigger other consequential actions without a separate authorization boundary.

Content

No source excerpt is available for this finding.

State-changing exploratory tests lack a warning about modifying real application data.

Medium
Category
Output Handling
Confidence
95% confidence
Finding

The guide directs open-ended form submissions, rapid repeated submissions, and record creation against an arbitrary target without establishing a disposable-data boundary. A clean browser context does not isolate backend state, so these actions can still modify live records or trigger external effects.

Content

No source excerpt is available for this finding.

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
75% confidence
Finding

Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.

Content

Scanner excerpt · references/exploratory-testing.md (reported line 75)May include surrounding context.

示例:

text
FINDING: Settings page — "Regenerate API Key" button has no confirmation dialog
SEVERITY: high
ROUTE: /orgs/:slug/:projectId/settings/general
EVIDENCE: Clicked "Regenerate" and the key changed immediately with no warning