Back to skill

Security audit

AI替你干网页活·登录一次就够

Security checks for vulnerabilities and agentic risk

Overview

ClawScan rates this browser-profile automation skill suspicious with high confidence because scanners report serious cookie disclosure, browser-process termination, and account-binding risks; the scan is partial, so its evidence is incomplete.

Consider avoiding installation until the cookie replay, browser-process cleanup, and account-binding risks are addressed or independently reviewed. The scan is partial: only half of the listed components were scanned, several static analyzers were degraded, and some referenced paths were missing, so it cannot establish that unreported behavior is safe.

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 (5)

T09 · Insecure Skill Coding Practices

Error
Location
scripts/supervisor.js
Finding

Cross-origin requests can invoke the privileged loopback API

Content
View full analysis

Impact

An attacker can execute JavaScript in a logged-in page or cause a live slot's cookies to be sent to an attacker endpoint through replay. Reading the API response is unnecessary for these effects.

T09 · Insecure Skill Coding Practices

Error
Location
scripts/supervisor.js
Finding

Replay sends every domain's cookies to a single destination

Content
View full analysis

Impact

The destination can obtain unrelated session cookies, including HttpOnly cookies accessible through CDP, potentially enabling account takeover. HTTP destinations also receive Secure cookies without TLS.

T09 · Insecure Skill Coding Practices

Error
Location
scripts/scripter.js
Finding

Websites can forge recorder commands and cross-origin playback steps

Content
View full analysis

Impact

A saved workflow can contain actions the user never recorded, including clicks on sensitive controls at another logged-in origin when the user later runs the workflow.

T09 · Insecure Skill Coding Practices

Warning
Location
scripts/scripter.js
Finding

Imported script identifiers escape the script storage directory

Content
View full analysis

Impact

An attacker can create or overwrite JSON files outside script storage, subject to the process's filesystem permissions and the imposed filename suffix. Native code execution is not established.

T09 · Insecure Skill Coding Practices

Warning
Location
scripts/gui.html
Finding

Imported identifiers inject JavaScript into GUI event handlers

Content
View full analysis

Impact

Injected handler code can call the same-origin farm API to access logged-in browser pages, issue authenticated actions, or disclose account data.

Vulnerability Patterns
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Output HandlingUnvalidated Output Injection, Cross-Context Output, Unbounded Output
  • YARA SignaturesMalware Match, Webshell Match, Cryptominer Match
  • MCP Least PrivilegeUnderdeclared Capability, Wildcard Permission, Missing Permission Declaration
Findings (12)

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

The automatic default-profile cleanup described in this source supports the concern about interfering with non-farm browser processes, although it does not establish indiscriminate system-wide termination. The separate claim that replay forwards every profile cookie cannot be confirmed from this document: authenticated replay is documented, but cookie-selection implementation is absent.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
90% confidence
Finding

The instructions disclose automatic cleanup of headless Chrome processes occupying the default profile, extending management beyond the farm's own isolated profiles. Legitimate unrelated automation could be terminated, interrupting work or losing unsaved state; the explicitly documented, user-selected Forge integration is not itself evidence of malicious or unexpected execution.

Content

No source excerpt is available for this finding.

Credential Access

High
Category
Privilege Escalation
Confidence
70% confidence
Finding

Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.

Content

Scanner excerpt · scripts/supervisor.js (reported line 276)May include surrounding context.

js
'--disable-client-side-phishing-detection',
    '--disable-domain-reliability',
    '--disable-breakpad',
    '--password-store=basic', '--use-mock-keychain',
    // 磁盘缓存默认无上限,多档一起吃磁盘:普通档锁 32MB,效率档砍到 8MB(够缓存表头脚本,不攒垃圾)
    '--disk-cache-size=' + (mode === 'eff' ? 8388608 : 33554432),
    // Translate/MediaRouter/喂流是纯用不上的功能;BackForwardCache+CalculateNativeWinOcclusion 关掉:

The default-enabled Chrome guard terminates browser processes outside the farm.

High
Category
Excessive Agency
Confidence
99% confidence
Finding

The default-enabled guard forcefully terminates headless Chrome process trees based on command-line heuristics rather than verified farm ownership. Legitimate browsers using relative or default profile directories can be killed, disrupting unrelated automation and potentially losing unsaved state; managing farm browsers does not authorize terminating other applications.

Content

No source excerpt is available for this finding.

Request replay forwards every profile cookie to the caller-selected destination.

High
Category
Excessive Agency
Confidence
100% confidence
Finding

Replay collects every cookie in the profile and sends the combined Cookie header to an arbitrary caller-selected URL, ignoring domain, path, and Secure restrictions. An attacker-controlled destination can therefore receive unrelated session credentials, and HTTP destinations can expose them in transit; persistent authenticated profiles make this especially dangerous.

Content

No source excerpt is available for this finding.

YARA rule 'info_stealer': Information stealer patterns (credential harvesting, browser data theft) [malware]

High
Category
YARA Match
Confidence
99% confidence
Finding

Cookie retrieval is legitimate for this skill's replay feature, so the match does not by itself demonstrate information-stealing malware. However, the surrounding implementation forwards all retrieved cookies to a caller-selected destination, confirming a credential-disclosure vulnerability that overlaps the replay finding.

Content

Scanner excerpt · scripts/supervisor.js (reported line 619)May include surrounding context.

js
log('档位 ' + slot.id + ' 停止抓包:' + requests.length + ' 个接口');
  return { ok: true, seconds: Math.round((Date.now() - s.startedAt) / 1000), count: requests.length, requests };
}

async function slotCookieHeader(slot) {
  if (!(await isAlive(slot))) throw new Error('档位没在跑,cookie 拿不到');
  const v = await jget(slot.port, '/json/version', 2000);
  // Network.getAllCookies 已被新版 Chrome 移除,cookie 走 Storage 域
  const r = await cdp(v.webSocketDebuggerUrl, 'Storage.getCookies');
  return r.cookies.map(c => c.name + '=' + c.value).join('; ');
}

async function replaySlot(slot, body) {
  // 带着档位登录态脱离浏览器直发——成功就证明这接口能转纯后台
  if (!body || !body.url) throw new Error('缺 url');
  const u = new URL(body.url);
  const mod = u.protocol === 'https:' ? require('https') : http;
  const cookie = await slotCookieHeader(slot);
  const headers = { ...(body.headers || {}), Cookie: cookie, 'User-Agent':

Configuration import can retain a script binding to an unrelated local account despite its wrong-account safeguard comment.

High
Category
Excessive Agency
Confidence
99% confidence
Finding

When an imported binding lacks an explicit old-to-new mapping, the importer accepts an existing local slot with the same numeric ID. Slot numbers do not establish account identity, so an imported script can remain bound to an unrelated authenticated account and perform unintended actions when started.

Content

No source excerpt is available for this finding.

Skill declares no tool scope ('permissions' or 'allowed-tools') but code capabilities were detected: env, network.

Medium
Category
MCP Least Privilege
Confidence
70% confidence
Finding

Without declared permissions the skill's intent is opaque and cannot be validated.

Content

No source excerpt is available for this finding.

Dangerous remote operations are presented as recoverable through snapshots without a warning about irreversible server-side changes.

Medium
Category
Output Handling
Confidence
98% confidence
Finding

The rule presents snapshots as a way to recover from dangerous bulk edits, deletions, and submissions without distinguishing local profile state from remote application state. An agent could authorize destructive authenticated actions under a false assumption of reversibility, while restoring browser files cannot undo server-side changes or financial transactions.

Content

No source excerpt is available for this finding.

Recursively deletes the test-home3 directory without disclosing the deletion.

Medium
Category
Output Handling
Confidence
94% confidence
Finding

The script unconditionally removes the entire test-home3 directory before running, including any pre-existing contents, and repeats the deletion on success and failure at lines 120 and 126. No prompt, log, or explanatory comment warns users about this destructive filesystem operation; the introductory comments describe only the idle-timeout test.

Content

No source excerpt is available for this finding.

The build command saves automation scripts even when its required trial validation fails.

Low
Category
Excessive Agency
Confidence
99% confidence
Finding

The build command saves the watch script after a failed trial or an evaluation exception, despite the comment saying validation must succeed first. Although saving does not immediately execute it, this permits unvalidated automation to be run later against authenticated dashboards, potentially clicking an unintended control or failing to enforce a spending threshold.

Content

No source excerpt is available for this finding.

The test claims to be offline but deliberately fetches a non-loopback hostname without a local mapping in the control case (L168–L175).

Low
Category
Excessive Agency
Confidence
94% confidence
Finding

L001 and L096 explicitly describe the test as offline and not touching external networks. However, the ordinary-browser control executes a fetch to sub.blocked.example before the efficiency-mode loopback mapping is enabled; a nonexistent or reserved domain does not prevent outbound DNS resolution attempts, so expected fetch failure does not establish offline operation.

Content

No source excerpt is available for this finding.