Monday, November 16, 2026 Edition No. 240
Search
Daily News Fully automated · 4 min read

AI Security News: November 11, 2026

Fully automated. Found, summarized and published by an automated AI pipeline. How posts are labeled

1. GPUStack Fixes a Critical SAML Authentication Bypass and a Kubernetes Proxy Authorization Gap, After CISA Lists an Earlier Loopback Bypass as Exploited

GPUStack, an open-source GPU cluster manager used to serve AI models, published two advisories on November 10, both fixed in version 2.2.2. CVE-2026-102894 (critical, CVSS 9.1) affects versions 0.7.1 to before 2.2.2: the /auth/saml/callback endpoint decodes a posted SAML response and issues a gpustack_session token without calling process_response(), the step that verifies signature, issuer, audience and time conditions. No credentials are needed and SAML SSO does not have to be configured, so any deployment reachable over the network is exposed; naming an existing administrator in the forged response returns that administrator’s session, and captured signed responses can be replayed indefinitely. CVE-2026-103002 (high, CVSS 8.8) affects versions 2.2.0 to before 2.2.2: the Kubernetes API proxy forwarded requests through a worker’s ServiceAccount without a resource-level authorization check, so an authenticated non-admin user could reach clusters they were not authorized for and, depending on the ServiceAccount’s permissions, read resources or create workloads. The disclosures follow CVE-2026-102895 (critical, CVSS 9.8), an unauthenticated administrator access flaw through loopback-source trust fixed in 2.2.0 that CISA added to its Known Exploited Vulnerabilities catalog on November 9. CVE-2026-102894 CVE-2026-103002

2. Spellwarden Census of the Official MCP Registry Finds 62% of Responding Hosts Listing Tools to Anonymous Clients

Spellwarden published a census of the official MCP Registry, last updated November 10, covering 59,944 active server entries and 30,402 unique hosts, with one endpoint probed per host using at most nine unauthenticated requests and no tool calls or credentials. Of 26,548 hosts that answered as MCP servers or requested sign-in, 16,470 (62%) listed tools to an anonymous client, and 9,691 of those open lists (59%) include at least one tool that does not declare itself read-only; at least 15,064 tools in open lists declare themselves destructive. The scanner’s instruction-like tool text rule fired on 967 hosts, and a manual review of a 60-host sample put the estimate at about 145 hosts steering the model toward their tools and about 96 asking it to hide something, follow a script or collect data. Among 25,455 local-package entries, 9,916 ask for at least one secret in an environment variable and only 17 of 1,729 container images are pinned by digest; 7,258 remote entries tell clients to send an authorization or API key header. The report states that it describes exposure to anonymous clients, not confirmed exploitation. Source

3. Kong Advisories Cover an AI Gateway Content-Safety Bypass for Large Prompts and Cross-Caller Cache Leaks

Kong published a batch of high-severity advisories on November 11 for Kong Gateway Enterprise and Kong AI Gateway. CVE-2026-96437 states that the ai-azure-content-safety plugin skips inspection of request bodies larger than the in-memory body buffer, allowing prompts to bypass input content-safety checks. CVE-2026-92443 states that the caching plugins may serve a response cached for a request carrying an Authorization or Proxy-Authorization header to different callers, and CVE-2026-92545 states that the proxy-cache and proxy-cache-advanced plugins omit the request hostname from the cache key, which may allow a response cached for one hostname to be served for another. The advisory summaries do not list patched versions. CVE-2026-96437 CVE-2026-92443