Disclosed vulnerability
MCP: untrusted servers and confused clients - tool metadata as a prompt injection channel
AIR-2025-0032 AIR record · 2025-05-mcp-untrusted-servers-confused-clients
From the supporting source
Tool calling has its inherent dangers, regardless of the implementation details, be it OpenAPI, AI Actions or MCP - they all suffer from prompt injection and confused deputy threats.embracethered.com
The record
Disclosed
2025-05-02
Vendor
multiple
Product
Claude Desktop, GitHub Copilot agent mode, Cursor, ChatGPT AI Actions
Risk domain
Agent behaviour › Unsafe actuation
Also touches
Data exfiltration channels, Confused deputy, Supply chain vulnerabilities
System type
Generative model
Who acted
Agent acted — tool metadata hijacked the client's conversation and tool calls
Attack surface
MCP server
Autonomy level
Semi-autonomous
Initial vector
Malicious tool description
Impact type
Unauthorized action, Data exfiltration
Confidence
High
Submitted by
EAI Internal
Crosswalk
- OWASP Agentic Top 10 (2026)
- ASI01 – Agent Goal Hijack
- MITRE ATLAS
- AML.T0048 – External Harms
- NIST AI RMF
- MEASURE 2.6; MANAGE 1.3
- EU AI Act
- Annex III §1(c)
- ISO/IEC
- TR 24028; 24029-1; 23894
- AIUC-1
- C. Safety
- Scope
- agent misuse
- Mitigation
- Require confirmation for destructive actions; Implement dry-run modes; Use layered safety controls.
- Taxonomy
- Enkrypt AI agent risk taxonomy · Black-Box Red Teaming of Agentic AI: A Taxonomy-Driven Framework for Automated Risk Discovery
Blast radius
any client connecting to an untrusted MCP server
Capability abused
Defense & remediation
Recorded defense
tool approval UI (does not surface hidden Unicode in tool descriptions)
Guardrail kinds
Human approval
Guardrail outcome
Silent
Fix shipped
no fix; disclosed to Anthropic 14+ months earlier and not treated as a security vulnerability
Taxonomy & references
Curator notes
Establishes that a malicious MCP server can hijack a conversation even without its tools being invoked.
Supporting sources
Cite as
Enkrypt AI. "MCP: untrusted servers and confused clients - tool metadata as a prompt injection channel". AIR, AIR-2025-0032. https://agentincident.vercel.app/i/AIR-2025-0032/ (accessed 2026-10-19).