Research demo
Malicious MCP tool descriptions can exfiltrate a user's entire conversation history on a trigger word
AIR-2025-0029 AIR record · 2025-04-trailofbits-mcp-conversation-theft
From the supporting source
Conversation histories are a gold mine of sensitive information, and this type of attack can be persistent: unlike traditional point-in-time data breaches, a single, sustained connection to a malicious MCP server could continue to compromise and harvest weeks or months of conversations.blog.trailofbits.com
The record
Disclosed
2025-04-23
Vendor
multiple
Product
MCP clients generally
Risk domain
Privacy › Data exfiltration channels
Also touches
Supply chain vulnerabilities
System type
Generative model
Who acted
Agent acted — the agent exfiltrated the conversation history on a trigger word
Attack surface
MCP server
Autonomy level
Copilot with approval
Initial vector
Malicious tool description
Impact type
Data exfiltration
Confidence
High
Submitted by
EAI Internal
Crosswalk
- OWASP Agentic Top 10 (2026)
- ASI02 – Tool Misuse and Exploitation
- MITRE ATLAS
- AML.T0024 – Exfiltration via AI Inference API
- NIST AI RMF
- MAP 4.2
- EU AI Act
- Article 10
- ISO/IEC
- TR 24028; 23894
- AIUC-1
- A. Data & Privacy
- Scope
- tool misuse
- Mitigation
- Monitor network traffic for anomalous patterns; Restrict egress destinations.
- Taxonomy
- Enkrypt AI agent risk taxonomy · Black-Box Red Teaming of Agentic AI: A Taxonomy-Driven Framework for Automated Risk Discovery
Blast radius
Any environment where a malicious MCP server is installed
Capability abused
Taxonomy & references
Curator notes
Author Keith Hoodlet. Second post in the Trail of Bits MCP line-jumping series.
Supporting sources
Cite as
Enkrypt AI. "Malicious MCP tool descriptions can exfiltrate a user's entire conversation history on a trigger word". AIR, AIR-2025-0029. https://agentincident.vercel.app/i/AIR-2025-0029/ (accessed 2026-10-19).