Skip to content
KitploitKITPLOIT
ToolsExploitsBlog
Log in
Submit
ToolsExploitsBlog
Submit

Hacking, PenTest, and Cybersecurity Tools for Your Security Arsenal!

Kitploit is a directory of hacking, cybersecurity, and pentesting tools. Discover the latest project updates to find vulnerabilities, analyze systems, automate testing, and strengthen your security.

FeedsContactPrivacy© 2026 Kitploit

Tool Directory

Categories

View all categories
Loading categories
Tools/GitHubGitHub/sustilliano/cve-2026-93328-gate
Authentication & AuthorizationDefensive ToolsVulnerability AnalysisExploitationCloud Security
GitHubsustilliano/cve-2026-93328-gate

cve-2026-93328-gate

The authorization layer CVE-2026-93328 was missing: scoped grants, path confinement, and signed receipts in front of MCP storage tools. CVE reproduction + gated proof, 10/10 tests.

View Repository
1423 days agoNot yet reviewed
Share

The layer CVE-2026-93328 was missing

CVE-2026-93328 / GHSA-2gc9-c85v-5vwr — Path Traversal and Missing Authorization in the cloud-storage source of Google Cloud MCP Toolbox for Databases (googleapis/mcp-toolbox), versions 1.2.0 through 1.4.0. CVSS 9.3 (Critical). CWE-22 (path traversal) + CWE-862 (missing authorization). Published 2026-10-26. Per the advisory, an unauthenticated remote attacker can exfiltrate host files to external Google Cloud Storage buckets and overwrite arbitrary local files via crafted tool invocation parameters.

This repository is the authorization layer that was missing — built, tested, and published so anyone can run it and prove or disprove the claim.

What the vulnerable code actually does (read at tag v1.4.0)

The cloud-storage tools take bucket, object, and the local path entirely from caller-supplied tool parameters:

  • cloudstorageuploadobject's invoke path passes the caller's localPath through ValidateLocalPath, then the source opens that file (os.Open) and uploads it to the caller-chosen bucket.
  • cloudstoragedownloadobject mirrors it: the source creates the caller-chosen destination (os.OpenFile, mode 0644) and writes the bucket object's bytes there.

ValidateLocalPath (in internal/tools/cloudstorage/cloudstoragecommon/paths.go) enforces a shape contract only — non-empty, absolute after cleaning, no .. segments. Its own comment says the quiet part out loud: OS permissions are the real boundary; the check just prevents "obvious traversal mistakes." It confines the path to nothing. Every absolute path on the host — /etc/passwd, a user's SSH keys, application secrets — passes it. And there is no caller authorization in front of the tools at all: the request executes with the server process's ambient filesystem authority. That combination is the CVE: a validation function mistaken for an authorization boundary, in tooling whose inputs are chosen by AI agents — callers that are supposed to be untrusted.

What upstream did

Google's remediation (PR googleapis/mcp-toolbox#3454, merged 2026-06-18, reported by HE WEI) added allowedBuckets and allowedLocalRoot options to the cloud-storage source. The PR description states the default plainly: "Defaults to allow all."

If you run mcp-toolbox: upgrade past the affected range and set the allowlists — the default preserves the vulnerable posture.

That default is the pattern this repository argues against. Protection that ships turned off, in the exact component whose job is standing between an agent and the host filesystem, is not a gate. It is a setting — and settings get left unset.

What this crate does

mcp-authz-gate puts the same two operations (upload_object, download_object, over buckets modeled as local directories so the proof needs no cloud account) behind four mechanisms:

  1. Scoped grants. The operator issues a per-caller key whose grant names its local root, its buckets, and its permitted operations. The raw key is shown once; the gate keeps only its SHA-256 hash.
  2. Confinement by resolution. Every path is resolved against the live filesystem — symlinks followed, deepest existing ancestor canonicalized — and must land inside the grant's root. A path that merely looks inside does not pass.
  3. Fail closed. No key, unknown key, revoked grant, out-of-scope bucket, escaping path: denied. There is no allow-all mode.
  4. Signed receipts. Every invocation — allowed or denied — is appended to a hash-chained journal signed with Ed25519, verifiable by anyone holding the public key. Attempts become evidence, not log lines that can quietly disappear.

Prove it / disprove it

cargo test

The suite (13 tests) runs both shapes against the same sandbox:

  • The vulnerable shape reproduces the CVE: an unauthenticated upload reads a file outside any intended root into an attacker bucket; a download overwrites an arbitrary host file; and a dedicated test documents that the v1.4.0 shape validator passes /etc/passwd and ~/.ssh/id_rsa — shape, not confinement.
  • The gated shape denies the identical calls — no key, forged key, wrong bucket, .. traversal, symlink escape, revoked grant — receipts each denial, lets legitimate in-scope work through, and the receipt chain verifies end to end (and fails under the wrong key).

Receipts for this build: 13 passed, 0 failed on 2026-10-26, from a clean checkout of this repository's own tree.

Revision note (v0.1.1, 2026-10-26). The first published version of this crate was reviewed and found wanting in three places — the same class of gap it exists to criticize, so they are recorded here rather than quietly patched. (1) The bucket/object join in the local simulation was unchecked, letting a crafted object name escape the simulated bucket; object names are now confined (in real GCS they are keys, not paths — the escape was an artifact of the stand-in, and artifacts count). (2) The path resolver treated a dangling symlink destination as a nonexistent filename and returned a path whose write would have followed the link outside the grant root; unresolvable symlink components are now rejected, matching the treatment upstream later adopted. (3) Several post-authorization I/O failures returned before a receipt was appended; every invocation now appends exactly one receipt — allowed, denied, or failed. Each defect has a regression test in the suite. The broader lesson is the repository's own thesis: passing tests establish only the guarantees they exercise.

Scope, stated plainly

This is a demonstration layer, not a patch for Google's binary: mcp-toolbox is a Go server, this crate is Rust, and buckets here are local directories standing in for GCS. What transfers is the architecture — scoped grants, resolution-based confinement, fail-closed defaults, and signed witness receipts in front of agent-invoked tools. What does not transfer is any claim about Google's current release: read their code and their advisory, not this README, for that.

By the PlaneKey project — where keys are scoped, custody is hashed, revocation is immediate and journaled, and no tool call happens without a receipt.

Download Tool

Most Popular

View all →

Discover the most used tools by our community.

  • 1

    OsintHunter

    letox4/osinthunter

    🎯 The ultimate OSINT toolkit: investigation graph, confidence scoring, alias families, map, timeline and reports. 49 modules, 400+ sites, no API keys required.

    795
  • 2

    ghostlock-app

    yukonga/ghostlock-app

    GhostLock One-Tap Execution App (CVE-2026-43499)

    520
  • 3

    ARTEX

    baarai/artex

    LLM multi-agent autonomous penetration testing platform with planner/worker agents, asset and exploration graphs, MITM traffic recording, and vulnerability retesting.

    439
  • 4

    redforge

    cybersecurity91/redforge

    Browser-side red team toolkit with 102 offline tools for recon, payload generation, credential attacks, crypto, encoding and engagement reporting.

    346
  • 5

    looking-glass

    anonghosty/looking-glass

    CSIRT platform that ingests Shadowserver reports into MongoDB and replays them as interactive maps, analysis, findings, advisories and executive dashboards.

    267
  • 6

    CyberStrikeAI

    ed1s0nz/cyberstrikeai

    The system of action for AI-native cybersecurity—where intent becomes governed execution, evidence becomes operational memory, and every operation improves the next.

    264

Explore all tools

Browse our collection of tools

View all tools →