Disclosed vulnerability
Prompt injection to host RCE in Microsoft Semantic Kernel's Search Plugin (eval() on a filter parameter)
AIR-2026-0106 AIR record · 2026-05-microsoft-semantic-kernel-rce
From the supporting source
We discovered a vulnerable path in Microsoft Semantic Kernel that could turn prompt injection into host-level remote code execution (RCE).microsoft.com
The record
Disclosed
2026-05-07
Vendor
Microsoft
Product
Microsoft Semantic Kernel (Search Plugin + In-Memory Vector Store)
Risk domain
Access control and permissions › Privilege escalation
System type
Generative model
Who acted
Agent acted (contested) — the plugin eval()'d a model-controlled filter parameter
Attack surface
Agent framework
Autonomy level
Semi-autonomous
Initial vector
Direct prompt injection user input
Impact type
Code execution, Remote code execution, Sandbox escape, Arbitrary file write
Confidence
High
Submitted by
EAI Internal
Crosswalk
- OWASP Agentic Top 10 (2026)
- ASI03 – Identity and Privilege Abuse
- MITRE ATLAS
- AML.T0055 – Unsecured Credentials
- NIST AI RMF
- GOVERN 6.1
- EU AI Act
- Article 13
- ISO/IEC
- TR 24028; 42001; 23894
- AIUC-1
- B. Security
- Scope
- tool misuse
- Mitigation
- Apply least-privilege principle; Regular policy audits.
- Taxonomy
- Enkrypt AI agent risk taxonomy · Black-Box Red Teaming of Agentic AI: A Taxonomy-Driven Framework for Automated Risk Discovery
Blast radius
Semantic Kernel agents (27k+ GitHub stars) using the affected plugin path
Capability abused
Defense & remediation
Recorded defense
cloud-hosted code-interpreter sandbox (.NET case); none for the Python eval() filter path
Guardrail kinds
Sandbox isolation
Guardrail outcome
Silent
Fix shipped
patched by Microsoft; CVE-2026-26030
Taxonomy & references
Curator notes
Microsoft Defender Security Research Team. Illustrates the general pattern: an agent framework passes model-controlled strings into a dangerous sink. Also reported under the titles: 'When prompts become shells: Semantic Kernel eval() and arbitrary file write turn prompt injection into host RCE (CVE-2026-26030, CVE-2026-25592)'; 'When prompts become shells: prompt-injection-to-RCE in Microsoft Semantic Kernel'.
Supporting sources
Cite as
Enkrypt AI. "Prompt injection to host RCE in Microsoft Semantic Kernel's Search Plugin (eval() on a filter parameter)". AIR, AIR-2026-0106. https://agentincident.vercel.app/i/AIR-2026-0106/ (accessed 2026-10-19).