Attack ChainAIDerived
Hypothetical attack flow derived from CVE metadata – not a confirmed exploit chain
-
AccessPlant symlink report.png in workspace root
-
DeliveryInvoke take_screenshot with report.txt path
-
ExploitRoots validation passes before extension rewrite
-
ExecutionensureExtension redirects sink to report.png symlink
-
ImpactWrite follows symlink outside configured root
Vulnerability Assessment
| Exploitation Conditions | Requires all of: chrome-devtools-mcp version 1.1.0 through 1.4.0 running on Linux with MCP roots configured to restrict file access to a workspace; a local attacker able to write into that workspace root to pre-plant a symlink whose name carries the tool's enforced output extension (e.g. report.png) and that targets a victim-writable file outside the root; and the MCP agent being driven to call an extension-enforcing output sink - demonstrated with take_screenshot (src/tools/screenshot.ts:216), and reachable via the same McpContext.saveFile pattern used by script, lighthouse, performance, and NetworkFormatter/McpResponse exports - with a caller path (e.g. report.txt) that passes roots validation before the extension rewrite. Limiting factors: no network vector (local only), the attacker needs prior local write access to stage the symlink, the tool must be invoked (user/agent interaction), and the written bytes are largely tool-generated (screenshot/report content) rather than fully attacker-chosen, bounding integrity impact. |
| Impact | Signals are consistent and point to a genuine but moderate, locally-gated issue rather than a mass-exploitation emergency. The provided CVSS 3.1 vector (AV:L/AC:L/PR:N/UI:R/S:U/C:L/I:L/A:N = 4.4, Medium) matches the description's 'local attacker' framing: exploitation is local, not network-facing, and requires user/agent interaction (invoking an output-producing tool). No EPSS score and no CISA KEV entry are present in the input, so there is no evidence of active or widespread exploitation; the vuln.today decision label is 'this_month' (not act_now/emergency), reinforcing a planned-remediation rather than drop-everything posture. Counterbalancing this, the GHSA advisory contains detailed, runnable proof-of-concept scripts (a self-contained lab PoC plus a full MCP stdio PoC against real Chrome), so a working exploit path is publicly documented, and this is explicitly an incomplete-fix variant of the same boundary class as CVE-2026-53766 - meaning the root-cause area has recurred. Net: a real, patch-worthy MCP roots sandbox-escape, but its impact is bounded by the need for local filesystem access to pre-plant a symlink and for the tool to be triggered, keeping it below network-RCE urgency. |
| Exploit ScenarioDerived | On a developer workstation running chrome-devtools-mcp (1.1.0-1.4.0) on Linux with MCP roots configured, a local attacker (or a prior foothold) pre-plants a symlink named with an enforced output extension, e.g. workspace/report.png, pointing at a victim-writable file outside the workspace root. When the MCP agent is driven to run an output tool such as take_screenshot with filePath 'workspace/report.txt' (format png), roots validation passes on the .txt path but ensureExtension rewrites the sink to report.png after validation, so the write follows the symlink and lands outside the configured root. The GHSA advisory publishes working PoC code demonstrating exactly this overwrite end-to-end. |
Get an e-mail when this CVE is exploited, patched or re-scored. Free account.
Alert meAnalysisAIv1 – Oct 29, 2026
chrome-devtools-mcp versions 1.1.0 through 1.4.0 on Linux can be induced to write tool output outside the MCP roots workspace boundary, giving a local attacker limited arbitrary read/write against files the victim user can write. The bypass works because the caller-supplied path is canonicalized and validated first, then rewritten to the tool's enforced extension (for example report.txt becomes report.png) and written without a second validation, so a pre-planted symlink at the final extension-rewritten path redirects the write outside the root; the trigger requires only local, unprivileged access (CVSS:3.1/AV:L/AC:L/PR:N/UI:R, 4.4 Medium) but depends on an agent being driven to invoke an extension-enforcing sink such as take_screenshot. The upstream GHSA-5hw5-4q87-m6x9 advisory includes four reproduction scripts, including a full MCP stdio PoC against a real Chrome instance; no CISA KEV entry or confirmed in-the-wild exploitation is reported, and the vendor-released fix is version 1.5.0.
RemediationAI
Vendor-released patch: upgrade chrome-devtools-mcp to 1.5.0 (from any 1.1.0-1.4.0 install), per the GHSA-5hw5-4q87-m6x9 advisory; the upstream project treats a validated-then-rewritten output path as unresolved until that release, so upgrading is the complete fix. No workaround is documented in the available sources, so until patched, address the prerequisites described for exploitation: restrict who can create, modify, or symlink files anywhere inside the configured MCP roots workspace (this removes the ability to pre-plant a report.png-style symlink that targets a file outside the root, but does not protect roots writable by untrusted local users), and avoid driving the agent to extension-enforcing output sinks (take_screenshot, script, lighthouse, performance, and the network/formatter response exports) with caller-supplied paths, since those all route through the same McpContext.saveFile ordering (cost: loss of screenshot and report export functionality). The decision signal for this issue is a scheduled patch cycle (this_month), not emergency response, and the fix should be tracked against the exact 1.5.0 release rather than the earlier 0.24.0-1.0.1 fix for CVE-2026-53766, which does not cover this variant.
Affected VersionsENISA
Affected Packages
Source: OSV (Open Source Vulnerabilities)
Affects your stack
Affects 1 package in euvd
Checking this CVE against your own stacks and servers is part of the Basic plan and up.
Technical ContextAI · show details
chrome-devtools-mcp is an npm-published Model Context Protocol server that exposes Chrome DevTools automation tools (screenshots, script execution, Lighthouse and performance captures, network/formatter exports) to an MCP client or agent, and it is expected to confine file access to the workspace roots declared through MCP roots. The flaw is a classic path-traversal and link-following ordering defect (CWE-22, CWE-59, as recorded in the upstream advisory) rather than a missing sanitizer: the validator in src/McpContext.ts already canonicalizes paths via fs.realpath and correctly rejects a final sink such as workspace/report.png whose realpath resolves to an out-of-root victim file. The vulnerable sequence at src/McpContext.ts:794-810 calls validatePath() on the client-supplied file path, then calls ensureExtension() from src/utils/files.ts:18-24, which strips the existing extension and substitutes the enforced one, and then calls fs.writeFile() on that newly generated path with no second validation. ToolHandler.ts:223-226 and src/tools/screenshot.ts:216 reach the same shared McpContext.saveFile sink, and the advisory lists the same pattern in src/tools/script.ts:163, lighthouse.ts:118, performance.ts:205, NetworkFormatter.ts:92 and :124, and McpResponse.ts:553, :581 and :748. The advisory describes this as an incomplete-fix variant of the same boundary class as CVE-2026-53766, with versions 0.24.0 through 1.0.1 belonging to that earlier symlink-canonicalization class and 1.1.0 being the first release affected by this post-canonicalization bypass. Integrity impact is bounded because the bytes written are largely tool-generated output rather than fully attacker-chosen.
Know when Chrome Devtools Mcp gets its next vulnerability
- Watch products and get an email when something new is published
- Check this CVE against your own stack and servers
- Exploit and PoC references, threat intelligence from 22 sources
No credit card · 7-day full trial
Severity by source
AV:L/AC:L/PR:N/UI:R/S:U/C:L/I:L/A:N
Local symlink pre-plant and a triggered output tool give AV:L/UI:R; no MCP authentication is needed (PR:N); write/read escapes the root but content is only partially attacker-controlled, so C:L/I:L/A:N.
Primary rating from Vendor (Chrome).
CVSS VectorVendor: Chrome
Lifecycle Timeline
3DescriptionCVE.org
Insufficient policy enforcement in chrome-devtools-mcp in Google Chrome on Linux prior to 1.5.0 allowed a local attacker to perform arbitrary read/write via a malicious file. (Chromium security severity: Medium)
Threat Intelligence Overview
How this decision was reached
The engine reads each signal below and applies a fixed list of rules from top to bottom. The first rule that matches sets the recommended action – like a statement, every line is shown.
What every action level means
SSVC AssessmentENISA
Priority Score
Attack Technique
Information disclosure occurs when an application unintentionally exposes sensitive data that aids attackers in reconnaissance or directly compromises security.
EUVD-2026-112366
GHSA-2mw5-2f2c-v2g3
NVD
CVE.org
Vulnrichment
VulDB
Exploit-DB