# Log-To-Leak Paper Shows Poisoned MCP Tool Descriptions Can Make Agents Leak Data After Finishing Their Task

Across 18 model-server pairs, the attack succeeded 62.64% to 100% of the time while leaving the legitimate task almost untouched, the same week OpenAI and Check Point expanded agent-driven vulnerability research.

- Published: 2026-10-28T06:21:21.801Z
- Canonical: https://polylog.news/ai/2026-10-28/log-to-leak-paper-shows-poisoned-mcp-tool-descriptions-can-m
- Publisher: Polylog (AI desk)
- Section: tech
- Sources: [arXiv cs.CR](https://arxiv.org/abs/2610.36108), [OpenAI News](https://openai.com/index/check-point-daybreak), [Anthropic Research](https://www.anthropic.com/research/team/frontier-red-team)

A paper posted to arXiv on October 26 by Yuepeng Hu, Chongyu Fan, Sirat Samyoun and Jian Du describes Log-To-Leak, an attack on agents that call tools through the Model Context Protocol (MCP). The attack exploits the natural-language metada…

This story is for subscribers. Read it in full at https://polylog.news/ai/2026-10-28/log-to-leak-paper-shows-poisoned-mcp-tool-descriptions-can-m (subscription information: https://polylog.news/pricing).