Friday, November 20, 2026 Edition No. 244
Search
Daily News Fully automated · 4 min read

AI Security News: November 19, 2026

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

1. mssql-mcp-core Fixes a High-Severity SQL Injection in Its MCP Write Tools That Bypassed Confirmation and Row-Limit Safeguards (CVE-2026-63129)

GitHub published an advisory for mssql-mcp-core, an npm package behind MCP servers for Microsoft SQL Server, rated high severity with a CVSS 3.1 score of 8.8 (CVE-2026-63129, CWE-89). Tools including update_data, delete_data, insert_data, execute_transaction, create_table and list_databases built T-SQL by interpolating caller-supplied table names, column names and a free-form whereClause. Because preview queries ran before the confirmation step, a crafted input executed on the first call and bypassed the preview, confirmation and maxRows controls, and with a privileged login and xp_cmdshell enabled it could reach OS command execution. The advisory says an LLM agent steered by indirect prompt injection could supply such arguments; the fix is in 0.6.0 for the core and server packages and 0.5.0 for the reader and writer packages. Source

2. LLMKube, a Kubernetes Operator for Running LLMs, Lets Namespace Users Read Arbitrary Host Files (CVE-2026-68903)

An advisory for LLMKube reports a high-severity path handling flaw (CVE-2026-68903, CVSS 3.1 score 7.7, CWE-22) in versions up to 0.8.28. The Model.spec.source field accepts absolute paths and file:// URIs without an allowlist, and passes them both to a file open in the controller pod and to a hostPath volume mounted into the inference pod. Any namespace user who can create Model and InferenceService resources could read files such as the controller’s service-account token or /etc/kubernetes/kubelet.conf on the node. Model caching, which activates the hostPath path, is enabled by default in the CLI and the official Helm chart, and the issue is fixed in 0.9.0. Source

3. Nudge Security Report Finds AI Agents Often Run on the Credentials of the Employee Who Built Them

Nudge Security released its AI Adoption in Practice 2026 report, based on anonymized telemetry from its customer environments, as covered by Help Net Security. Among organizations with at least one AI agent, the median count was seven, about 44% had more than 10, and agents ran across 15 platforms. Agents commonly run on the OAuth grant or API key of the employee who built them, or on a shared connection, so they inherit that person’s access. Anthropic integrations appeared in 74% of organizations and OpenAI in 67%, and pasted content or uploaded files made up 18% of observed prompt activity. The report recommends inventorying agents, reviewing OAuth grants given to AI tools and rotating any key or token that appears in a prompt; the article does not state the sample size. Source

4. Canonical Ships AI-Written Ubuntu Security Patches in an Opt-In Edge Archive Before Human Review Finishes

Canonical introduced an opt-in Edge archive for Ubuntu Pro customers that delivers security patches for userspace software written by Surge, an AI system the company built to fix known vulnerabilities. Per Help Net Security, agents match vulnerability reports to packages and gather the upstream fix, several models write competing backports, and a panel of judge models picks the best candidate, with uncertain cases sent to the Ubuntu security team. Approved patches go through the same deterministic tests as other Ubuntu packages, but they ship before engineers finish review and soak testing, and users who opt in accept responsibility for judging whether a patch is safe for production. Kernel fixes are excluded, and Canonical says Ubuntu’s security fixes have roughly doubled compared with last year. Source