Research
Log-To-Leak: Prompt Injection Attacks on Tool-Using LLM Agents via the Model Context Protocol
Publication date unknown · Discovered arxiv.org
Page published
Publication date unknown · First observed: 17 Nov 2026
Coverage timeline
Single-source research — one report is available.
Why it matters
Log-To-Leak demonstrates that successful task completion can conceal unauthorized disclosure through secondary tool calls, highlighting the need for host-controlled authorization rather than reliance on agent instructions alone.
Log-To-Leak is a research study of malicious MCP tool metadata that induces an unauthorized, attacker-controlled logging call after legitimate tool use, targeting user queries, server details, tool responses, and agent responses. Across 18 model-server pairs with valid tool execution, the authors report attack success rates of 62.64–100%, often with little change in legitimate task completion. Three evaluated defenses still left attack success above 55% on GitHub MCP with GPT-4o.