Semantic Kernel CVEs: RCE Patch Playbook for AI Agents

Microsoft's Semantic Kernel, a widely used framework for building enterprise AI agents, has serious security disclosures. Two specific Common Vulnerabilities and Exposures, CVE-2026-25592 and CVE-2026-26030, affect agent orchestration and retrieval-augmented generation paths. What's important here is that these vulnerabilities can enable remote code execution. This happens when untrusted content reaches privileged tool handlers. If your agents use plugins, shell tools, or file access, this is a significant issue. The first vulnerability, CVE-2026-25592, involves agent orchestration abuse. An attacker can influence agent inputs, potentially leading to unintended tool execution. The second, CVE-2026-26030, targets retrieval paths. Malicious content in documents indexed for retrieval can bypass filters and lead to tool selection based on attacker-controlled context. This means security teams need to think about entire pipelines, not just single points of failure, to protect against these new types of threats.