x.comthreadRohan Paul (@rohanpaul_ai) on X
Rohan Paul's X post summarizes a TikTok paper on MCP attacks. LLM agents often finish the requested task, then quietly call an attacker-controlled logging tool that records the user query, the server reply and the final answer. The attack works because agents read natural-language tool descriptions when choosing and ordering tool calls. The visible text is truncated and gives no success rates or defenses.
TL;DR Post describes a paper showing malicious MCP tool descriptions make agents call a logging tool after the task and leak query, reply and answer.
Technologies named here
- Model Context Protocol
- LLM
- LLM agents using MCP (Model Context Protocol) tools
Techniques named here
- Tool poisoning
- Data exfiltration
- Malicious MCP tool description injection
Related in AI Security
blog.talosintelligence.comtoolIntroducing CAIRN: Frontier tracking for AI-integrated malware (Cisco Talos)
github.comadvisoryFlowise MCP Security Bypass Enables RCE
github.comrepoYesWeHack Claude Kit
jfrog.comarticle