An authentication bypass vulnerability in the network driver of Palo Alto Networks Prisma® Access Agent on Windows enables a local administrator to bypass security inspection, subsequently allowing them to inject and intercept arbitrary network traffic.

The Prisma Access Agent on Linux, macOS, iOS, Android, and Chrome OS is not affected.

An improper link resolution before file access vulnerability exists in the Palo Alto Networks Prisma® Access Agent on Linux platforms that enables a local low privileged user to delete system files in a limited scope and disable Prisma Access Agent.

The Prisma Access Agent on macOS, Windows, iOS, Android, and Chrome OS is not affected.

Multiple Cross-Site Scripting (XSS) vulnerabilities were identified in GMS 9.5.1 (Build 9510.1044) and earlier versions that allow a remote attacker to execute javascript script in a user's browser.

Finding The statically-linked gRPC stack is pinned to v1.46.3 (2022) in the Bazel WORKSPACE file, predating multiple HTTP/2 DoS CVEs that are directly reachable on the network listener:

WORKSPACE:116-117 comgithubgrpcgrpc → gRPC 1.46.3 WORKSPACE:129-130 comgoogleprotobuf → protobuf 3.21.12 WORKSPACE:173-176 zlib → 1.3 The MLMD server is a network-facing gRPC listener (FIND-001: no auth), so gRPC-layer DoS CVEs are directly reachable from any pod that can open a TCP connection to :8080.

The Bazel WORKSPACE pins are content-addressed (sha256), so the issue is staleness, not mutability. Renovate is present (.github/renovate.json) but evidently not covering Bazel httparchive entries.

File: WORKSPACE:116-117,129-130,173-176 Repository: red-hat-data-services/ml-metadata Framework: ASVS V14.2.1; OWASP K8s K07 Vulnerable Components; OpenSSF Scorecard Vulnerabilities CWE: CWE-1395 / CWE-1104 CVSS v3.1: 7.5 (High) AV:N/AC:L/PR:N/UI:N/S:U/C:N/I:N/A:H (inherits gRPC HTTP/2 DoS vectors)

RHOAI Mitigation The DSPO-deployed NetworkPolicy restricts who can reach MLMD port 8080 to only KFP v2 driver pods and DSP components. This limits the attack surface but does not eliminate it — a compromise of a KFP driver pod or DSP component could exploit these CVEs to crash the MLMD pod and disrupt all pipeline runs in the namespace.

Impact An in-cluster attacker who can reach the MLMD pod (within the NetworkPolicy allowlist) can crash or resource-exhaust the MLMD pod via known gRPC/HTTP2 frame-handling bugs, disrupting all pipeline runs in the namespace.

Context ml-metadata is planned for removal from the product (several months out). The stale dependency risk remains active until removal is complete.

Remediation Bump WORKSPACE pins: gRPC >= 1.62, protobuf >= 25.x, zlib >= 1.3.1. Extend Renovate configuration to cover Bazel httparchive entries.

An allocation of resources without limits vulnerability in the HTTP handler component of Google mcp-toolbox versions up to and including 1.4.0 allows an unauthenticated attacker to cause a denial of service (DoS). The /mcp endpoint handler reads incoming payloads directly into system memory using an unrestricted buffer loop (io.ReadAll) without applying defensive constraints such as http.MaxBytesReader or pre-read Content-Length enforcement. By submitting a single, massive HTTP request body, an attacker can linearly consume available host memory until the runtime process is terminated by an Out-Of-Memory (OOM) error.

An improper authorization and security-boundary bypass vulnerability in the bigquery-execute-sql tool component of Google mcp-toolbox versions 0.16.1 through 1.4.0 allows an authenticated attacker to bypass allowedDatasets validation checks. The toolbox relies on the BigQuery dry-run API to enforce dataset restrictions, but due to a fail-open logic flaw, it bypasses validation when the API returns an empty array for specialized constructs. This allows the attacker to extract structural DDL schemas for explicitly excluded datasets via INFORMATIONSCHEMA, and access downstream federated row data via EXTERNALQUERY connections.

Information leak in Skia in Google Chrome prior to 151.0.7922.169 allowed a remote attacker to potentially bypass web origin policy via a crafted HTML page. (Chromium security severity: High)

CAI Content Credentials is affected by an Uncontrolled Resource Consumption vulnerability that could lead to application denial-of-service. An attacker could exploit this vulnerability to exhaust system resources, resulting in an application denial-of-service condition. Exploitation of this issue does not require user interaction.

CAI Content Credentials is affected by an Integer Underflow (Wrap or Wraparound) vulnerability that could result in an application denial-of-service. An attacker could exploit this vulnerability to crash the application, leading to a denial-of-service condition. Exploitation of this issue does not require user interaction.

CAI Content Credentials is affected by an Improper Input Validation vulnerability that could result in a Security feature bypass. An attacker could leverage this vulnerability to bypass security measures and gain unauthorized write access. Exploitation of this issue does not require user interaction.

CAI Content Credentials is affected by an Improper Input Validation vulnerability that could result in an application denial-of-service. An attacker could exploit this vulnerability to crash the application, leading to a denial-of-service condition. Exploitation of this issue does not require user interaction.

CAI Content Credentials is affected by an Integer Overflow or Wraparound vulnerability that could result in an application denial-of-service. An attacker could exploit this vulnerability to crash the application, leading to a denial-of-service condition. Exploitation of this issue does not require user interaction.

CAI Content Credentials is affected by an Integer Underflow (Wrap or Wraparound) vulnerability that could result in an application denial-of-service. An attacker could exploit this vulnerability to crash the application, leading to a denial-of-service condition. Exploitation of this issue does not require user interaction.

CAI Content Credentials is affected by an Improper Input Validation vulnerability that could lead to arbitrary file system read. An attacker could exploit this vulnerability to access sensitive files and directories outside the intended access scope. Exploitation of this issue requires user interaction in that a victim must open a malicious file.

The Widgets for Google Reviews plugin for WordPress is vulnerable to Stored Cross-Site Scripting via admin settings in all versions up to, and including, 13.3 due to insufficient input sanitization and output escaping. This makes it possible for authenticated attackers, with editor-level permissions and above, to inject arbitrary web scripts in pages that will execute whenever a user accesses an injected page. This only affects multi-site installations and installations where unfilteredhtml has been disabled.

An Improper Input Validation vulnerability in BigQuery DAO in Google Cloud Apigee versions prior to 2026-06-12 on Google Cloud Platform allows an authenticated attacker to exfiltrate cross-tenant data.

This vulnerability was patched on 12 June 2026 on the Apigee Servers, and no customer action is needed.

Chromium: CVE-2026-15131 Insufficient data validation in Navigation

Chromium: CVE-2026-15128 Inappropriate implementation in Forms

Chromium: CVE-2026-15130 Insufficient policy enforcement in Navigation

Chromium: CVE-2026-15127 Inappropriate implementation in WebGL

Chromium: CVE-2026-15124 Insufficient policy enforcement in Passwords

Chromium: CVE-2026-15109 Uninitialized Use in ANGLE

Chromium: CVE-2026-15108 Integer overflow in Extensions API

Handshakes which used Encrypted Client Hello could be de-anonymized by a passive network observer due to a disclosure of pre-shared key identities in the unencrypted client hello.

Chromium: CVE-2026-15775 Insufficient policy enforcement in V8

Microsoft Edge for Android Information Disclosure Vulnerability

Absolute path traversal in Microsoft Edge for Android allows an unauthorized attacker to disclose information locally.

Improper access control in Microsoft Edge for Android allows an unauthorized attacker to bypass a security feature over a network.

Chromium: CVE-2026-17897 Inappropriate implementation in ORB

Summary

I originally reported this through Google Bug Hunters. The Google Bug Hunters team said this is in OSS VRP scope but not reward-eligible due to the project tier, and asked me to file an issue or PR directly with this repository. I am reporting it privately here first because it is an unfixed security issue.

McpContext.validatePath() enforces workspace roots by checking whether path.resolve(filePath) textually falls under one of the configured root paths. path.resolve() does not canonicalize symbolic links. As a result, a symlink inside a configured workspace root can point to a file outside that root, pass validation, and then be followed by downstream file read/write operations.

This bypass applies even when the MCP client correctly declares the roots capability with a non-empty list. It is separate from the documented legacy behavior where missing roots capability allows all paths.

The practical impact is a workspace-boundary bypass. In the write direction, filePath-writing tools can overwrite out-of-root files through an in-root symlink. In the read direction, uploadfile can read through the symlink and send the file to the currently selected web page.

Details

Affected code:

src/McpContext.ts:178-199

ts validatePath(filePath?: string): void { if (filePath === undefined) { return; } const roots = this.roots(); if (roots === undefined) { return; } const absolutePath = path.resolve(filePath); for (const root of roots) { const rootPath = path.resolve(fileURLToPath(root.uri)); if ( absolutePath === rootPath || absolutePath.startsWith(rootPath + path.sep) ) { return; } } throw new Error( Access denied: path ${filePath} is not within any of the workspace roots ${JSON.stringify(roots)}., ); }

path.resolve() only normalizes path text such as . and ... It does not call realpath() and does not resolve symlinks. Therefore, a path like:

text /workspace/project/cache/profile

can textually pass the /workspace prefix check even when cache/profile is a symlink to:

text /home/user/.aws/credentials

Downstream consumers then perform real filesystem operations without ONOFOLLOW:

- src/McpContext.ts:720-738 saveFile() uses fs.mkdir({recursive: true}) and fs.writeFile(). - src/tools/input.ts:454-497 uploadfile calls puppeteer.uploadFile(filePath) or fileChooser.accept([filePath]). - Other filePath-writing tools include screenshots, heap snapshots, network response save paths, snapshots, screencasts, Lighthouse output, and performance trace saves.

This is not a TOCTOU/race condition. The symlink exists before validation and the PoC uses a single process. The issue is a canonicalization bypass / improper link resolution.

Preconditions:

- The MCP client declares roots and supplies at least one workspace root. - A symlink exists inside the workspace and points outside the workspace. - For the remote prompt-injection chain, the user processes untrusted page content while chrome-devtools-mcp is connected.

A remote attacker does not need local access if a suitable workspace-internal symlink already exists, or if another trusted tool/workflow can create it. Without such a symlink, the issue is a local/workspace-state-dependent boundary bypass.

PoC

Conceptual exploitation with a configured root:

text Configured roots: file:///workspace

Workspace path: /workspace/project/cache/profile -> /home/user/.aws/credentials

Tool call: uploadfile({ filePath: "/workspace/project/cache/profile", uid: "<file input element on current page>" })

Result: validatePath() accepts the path because it textually starts with /workspace. Puppeteer follows the symlink and uploads the target file to the page.

Lab-only PoC that replicates the exact validation logic and subsequent write. It writes only inside a fresh temporary directory and touches no system paths:

js const path = require('node:path'); const fs = require('node:fs'); const os = require('node:os'); const {pathToFileURL, fileURLToPath} = require('node:url');

const lab = fs.mkdtempSync(path.join(os.tmpdir(), 'cdtmcp-lab-'));

try { fs.chmodSync(lab, 0o755);

const workspace = path.join(lab, 'workspace'); fs.mkdirSync(workspace);

const outside = path.join(lab, 'outside-secret.txt'); fs.writeFileSync(outside, 'sensitive outside content\n');

const symlinkInside = path.join(workspace, 'innocent.txt'); fs.symlinkSync(outside, symlinkInside);

function validatePath(filePath, roots) { const absolutePath = path.resolve(filePath); for (const root of roots) { const rootPath = path.resolve(fileURLToPath(root.uri)); if ( absolutePath === rootPath || absolutePath.startsWith(rootPath + path.sep) ) { return true; } } throw new Error(Access denied: ${filePath}); }

const roots = [{uri: pathToFileURL(workspace).href, name: 'workspace'}]; validatePath(symlinkInside, roots);

fs.writeFileSync(symlinkInside, 'OVERWRITTEN BY MCP\n');

console.log(fs.readFileSync(outside, 'utf8')); // -> "OVERWRITTEN BY MCP" } finally { fs.rmSync(lab, {recursive: true, force: true}); }

Observed result:

text validatePath() accepts the in-root symlink path. The subsequent write follows the symlink and modifies the out-of-root target.

I can provide an end-to-end MCP client reproduction if needed. The lab PoC above demonstrates the root cause using the same validation logic as the server.

Impact

Who can exploit:

- A local process/user or trusted workflow that can create a symlink inside the workspace. - A remote page/prompt-injection attacker, if a suitable workspace-internal symlink already exists or can be created by another trusted workflow/tool.

Security impact:

- Integrity: tools that write to filePath can overwrite files outside the configured workspace root through an in-root symlink. - Confidentiality: uploadfile can read a file outside the workspace through an in-root symlink and attach it to a file input on the current page. - Stealth/auditability: the exfiltration path goes through normal page file-upload behavior, and chrome-devtools-mcp does not appear to log the canonical path that was uploaded.

Example sensitive files reachable if symlinked into the workspace:

- Cloud credentials such as ~/.aws/credentials, ~/.config/gcloud/..., or ~/.azure/.... - SSH private keys or .ssh files readable by the user. - Project secrets such as .env, .npmrc, .netrc, secrets.json, and API tokens. - Out-of-workspace source files or configuration files.

Severity:

- Suggested GitHub severity: Moderate. - CVSS v3.1 chain estimate: CVSS:3.1/AV:N/AC:H/PR:N/UI:R/S:U/C:H/I:N/A:N. - AC:H reflects that a workspace-internal symlink must exist at validation time.

Suggested fix:

Canonicalize paths before comparing against roots. For an existing file, use fs.realpath() on the path. For a new file, resolve the parent directory with fs.realpath() and re-join the basename.

ts async validatePath(filePath?: string): Promise<void> { if (filePath === undefined) return; const roots = this.roots(); if (roots === undefined) return;

const abs = path.resolve(filePath); let canonical; try { canonical = await fs.realpath(abs); } catch (err) { if (err.code === 'ENOENT') { const parent = await fs.realpath(path.dirname(abs)); canonical = path.join(parent, path.basename(abs)); } else { throw err; } }

for (const root of roots) { const canonicalRoot = await fs.realpath(fileURLToPath(root.uri)); if ( canonical === canonicalRoot || canonical.startsWith(canonicalRoot + path.sep) ) { return; } }

throw new Error( Access denied: ${filePath} (canonical: ${canonical}) is not within any workspace root., ); }