Skip to content
The Injection_
LIVE
LIVElast sweep next in--:--:--

[ RSS ][ JSON ][ llms.txt ]

Week 43 · October 19 to October 25, 2026

AI security this week

In week 43 of 2026 (October 19 to October 25, 2026), The Injection tracked 2 AI security stories, including "opencode: a remote MCP server could run commands on Windows" and "Google Agent Mesh: unauthenticated peers could invoke MCP tools". 1 story came with a CVE.

CVE-2026-103091
MEDVulnerability

Top storyMED

Google Agent Mesh: unauthenticated peers could invoke MCP tools

Two missing-authorization flaws let an outsider join Google's Agent Mesh or skip its policy check, then call tools on connected MCP servers.

What to do · Upgrade to v0.1.0-alpha.9 or later, which fails closed by default, requires explicit policy bindings at enrollment and adds ban checks. Replace any wildcard binding of sam:system:authenticated with explicit per-service bindings. Re-run dependency scanners against the github.com/google/sam Go module, because the CVE ids are new and earlier scans had nothing to match.

Medium 2

  1. MED

    Vulnerability · GitHub Advisory Database · CVE-2026-103091 · CVSS 9.3

    Google Agent Mesh: unauthenticated peers could invoke MCP tools

    Google Agent Mesh, an open-source mesh for MCP and A2A agents, received two critical CVEs on October 19, 2026 for flaws fixed in September: an unauthenticated peer could enroll itself and call tools on connected MCP backends.

  2. MED

    Vulnerability · Pillar Security

    opencode: a remote MCP server could run commands on Windows

    opencode ran attacker-supplied commands on Windows when a remote MCP server returned a crafted OAuth authorization endpoint. Anomaly fixed the flaw in 1.18.33 without a CVE, so scanners will not flag older installs.