Return to Threats

MCP Leaks, Poisoned Skill Files, and a Live Agent Incident

DEV Community 2026-11-08 data leakage Critical

Executive Summary

Research described a live agent attack involving injected MCP tool calls that could expose user queries and prior responses. Organizations operating AI agents or MCP-connected LLM applications should care because the activity may remain inconspicuous while leaking information.

What Happened

Research described attacker-controlled MCP tool calls that can be injected after an agent’s legitimate call, potentially exposing user queries and previous tool or agent responses. The testing reported success across multiple MCP servers and LLM agents without noticeably disrupting the requested task.

Affected Products

  • MCP servers
  • LLM agents

Vulnerability / Attack Technique

MCP tool-call injection

Technical Mechanism

Attacker-controlled MCP tool calls are injected after an agent's legitimate call. The reported testing indicated that the injected calls could expose user queries and previous tool or agent responses across multiple MCP servers and LLM agents without noticeably disrupting the requested task.

AI Security Category

LLM02:2025 Sensitive Information Disclosure

Part of the MCP Security topic.

Relevant Frameworks

LLM02:2025 Sensitive Information DisclosureLLM06:2025 Excessive AgencyMITRE ATLAS: Exfiltration

RealGround Analysis

This matters because MCP-connected agents may process sensitive user prompts and tool outputs while concealing malicious activity within an otherwise successful task. Teams should assess tool authorization, call sequencing, output exposure, and monitoring for anomalous MCP behavior.

Recommended Controls

  • Require explicit authorization and integrity validation for every MCP tool call, including calls occurring after the agent's primary action.
  • Restrict MCP tools to least-privilege access and prevent tools from reading unrelated user queries, prior agent responses, or tool outputs.
  • Log and monitor the complete MCP call sequence, including tool identity, parameters, returned data, and caller context.
  • Red-team MCP servers and LLM agents for post-call injection and covert data-exposure scenarios.
  • Apply output filtering and data-loss-prevention checks before tool or agent responses are returned to users or downstream systems.

Related RealGround Service

AI Agent Business Logic Audit · or start with the free AI Agent Security Scanner.

Sources

DEV Community: https://dev.to/felipe0liveira/the-channel-nobodys-watching-mcp-leaks-poisoned-skill-files-and-a-live-agent-incident-1m4j

Summary and analysis written by RealGround from the source above; last updated 2026-11-10.

Talk to AI CISO