Skip to main content
Vulnerability Database/CVE-2026-103091

CVE-2026-103091: Google OSS SAM Auth Bypass Vulnerability

CVE-2026-103091 is an authorization bypass flaw in Google OSS Agent Mesh that allows remote attackers to self-enroll and execute arbitrary commands across mesh nodes. This post covers the technical details, affected versions, impact, and mitigation steps.

Published:

Disclaimer: This content was generated using AI. While we strive for accuracy, please verify critical information with official sources.

Sources: National Vulnerability Database, CISA KEV, Exploit-DB

CVE-2026-103091 Overview

CVE-2026-103091 affects Google Open Source Software (OSS) Agent Mesh (SAM) versions 0.1.0-alpha.1 through 0.1.0-alpha.8. The vulnerability combines Missing Authorization [CWE-862] and Initialization of a Resource with an Insecure Default [CWE-1188] across the Model Context Protocol (MCP) proxy and the OpenID Connect (OIDC) registration flow. A remote attacker can self-enroll into the mesh and invoke arbitrary tools across every connected MCP backend node. The default policy binds the sam:system:authenticated group to wildcard services and targets, granting full read, write, and execution access on backend services once an attacker obtains any valid OIDC token.

Critical Impact

Unauthenticated remote attackers can enroll into the mesh and execute arbitrary tools across all connected MCP backends, resulting in full compromise of confidentiality, integrity, and availability of backend services.

Affected Products

  • Google OSS Agent Mesh (SAM) 0.1.0-alpha.1
  • Google OSS Agent Mesh (SAM) 0.1.0-alpha.2 through 0.1.0-alpha.7
  • Google OSS Agent Mesh (SAM) 0.1.0-alpha.8

Discovery Timeline

  • 2026-10-19 - CVE-2026-103091 published to the National Vulnerability Database (NVD)
  • 2026-10-19 - Last updated in NVD database

Technical Details for CVE-2026-103091

Vulnerability Analysis

The vulnerability chains three design defects in the SAM control plane and MCP proxy. The GET /info endpoint in internal/controlplane/server.go publicly discloses OIDC configuration, allowing attackers to discover the trusted identity provider. The POST /register endpoint auto-enrolls any caller presenting a valid OIDC token without applying claim filtering, allowlisting, or ban checks. Once enrolled, the default authorization policy binds the sam:system:authenticated principal group to wildcard services and targets, granting unrestricted access to every registered MCP backend.

The HandleStreamPassThrough function in mcp_service.go then relays attacker-controlled MCP requests to backend nodes without authorization enforcement at the proxy layer. The chain converts a public token issuer into a mesh-wide execution primitive.

Root Cause

The root cause is an insecure default authorization policy combined with an open enrollment endpoint. SAM ships with a fail-open policy that treats any authenticated principal as fully trusted. No explicit policy binding is required at enrollment, and no mechanism filters tokens by issuer, audience, or custom claims. The MCP proxy inherits this trust decision rather than performing independent authorization at the tool-invocation boundary.

Attack Vector

An attacker queries GET /info to retrieve OIDC metadata, obtains a token from any accepted identity provider, and submits it to POST /register. The control plane returns mesh credentials bound to the wildcard default policy. The attacker then issues MCP tool-invocation calls through HandleStreamPassThrough, which forwards them to every connected backend node without further checks. Information disclosure [CWE-200] through GET /info provides the discovery primitive that makes the chain reliably exploitable over the network.

See the GitHub Security Advisory GHSA-cgmg-9xf6-rrgw for additional technical detail.

Detect Identity Attacks the Moment Authentication Fails

Singularity™ Identity surfaces credential misuse, lateral movement, and privilege escalation in real time, so you can contain the breach before it spreads.

Stop Identity Attacks

Detection Methods for CVE-2026-103091

Indicators of Compromise

  • Unexpected POST /register requests on the SAM control plane from unknown source addresses or user agents.
  • Mesh credential issuance events bound to the sam:system:authenticated principal without an associated operator approval record.
  • MCP tool invocations through HandleStreamPassThrough targeting multiple backend nodes from a single newly enrolled identity.
  • Access logs showing GET /info requests from external networks followed by registration attempts.

Detection Strategies

  • Audit control plane logs for POST /register calls whose OIDC token claims do not match the operator allowlist.
  • Correlate newly enrolled mesh identities with subsequent MCP tool-invocation volume, flagging accounts that reach wildcard services within minutes of enrollment.
  • Alert on any policy binding that resolves to wildcard services or targets for the sam:system:authenticated group.

Monitoring Recommendations

  • Forward SAM control plane and MCP proxy logs to a centralized analytics pipeline with retention sufficient for retrospective hunting.
  • Monitor outbound tool invocations per enrolled identity and baseline normal behavior to detect lateral access across backend nodes.
  • Track OIDC token issuer and audience values seen at POST /register and alert on unexpected issuers.

How to Mitigate CVE-2026-103091

Immediate Actions Required

  • Upgrade Google OSS Agent Mesh (SAM) to version 0.1.0-alpha.9 or later, which enforces a fail-closed default policy, explicit policy-binding requirements at enrollment, and ban checks.
  • Revoke all mesh credentials issued by vulnerable versions and re-enroll operators under the hardened policy.
  • Restrict network exposure of the SAM control plane so that GET /info and POST /register are not reachable from untrusted networks.
  • Audit backend MCP nodes for unauthorized tool invocations during the exposure window.

Patch Information

The fix is available in Google SAM release v0.1.0-alpha.9. The release enforces a fail-closed default authorization policy, requires explicit policy bindings on enrollment, and adds ban checks against the registration flow. Review the AgentMesh advisory GHSA-cgmg-9xf6-rrgw for upgrade guidance.

Workarounds

  • Place the SAM control plane behind an authenticating reverse proxy that enforces an allowlist of operator identities before requests reach POST /register.
  • Configure network policy to block external access to GET /info until the upgrade is complete.
  • Replace the default wildcard policy binding with explicit, least-privilege bindings scoped to specific services and targets.
  • Disable auto-enrollment and require manual operator approval for new mesh identities where supported.
bash
# Configuration example
# Upgrade SAM to the fixed release
git fetch --tags
git checkout v0.1.0-alpha.9

# Network-level containment until upgrade completes
iptables -A INPUT -p tcp --dport 8080 -s 10.0.0.0/8 -j ACCEPT
iptables -A INPUT -p tcp --dport 8080 -j DROP

Disclaimer: This content was generated using AI. While we strive for accuracy, please verify critical information with official sources.

Sources: National Vulnerability Database, CISA KEV, Exploit-DB

Experience the Most Advanced Cybersecurity Platform

See how the world’s most intelligent, autonomous cybersecurity platform can protect your organization today and into the future.