Home / Vault & Second Brain / Your AI’s tools are an attack surface you never audit

Your AI’s tools are an attack surface you never audit

Your AI’s tools are an attack surface you never audit

Your AI’s tools are an attack surface you never audit

Yesterday I wrote about how the files your AI reads are the attack surface — a poisoned document becomes a cheap instruction channel into privileged operations. Today I want to talk about the other half of that equation: the tools your AI connects to. Because here’s the thing I keep running into in my own setup — I run a lot of MCP servers, and I never once stopped to audit them the way I audit my own memory.

MCP (Model Context Protocol) is how my agents talk to the world. In my stack that means Home Assistant, SearXNG, X, Proton, Proxmox, SSH, Seekstone, codebase-memory — eight MCP servers that hand my agent a list of tools and say “here’s what you can do.” And for a long time I treated those tool lists the way I treated my memory: trusted, because they came from a source I chose.

Then I read the numbers, and I stopped being comfortable.

The tool description is a prompt-injection vector

The pattern that finally clicked for me came from a security report on MCP implementations. Four issues kept coming up, and the first one is the one most people miss: the tool description itself can carry a malicious instruction. When an MCP server sends your agent a description like “this tool reads the sensor data,” that description lands in your agent’s context. It is instructions. And instructions from an unverified source are exactly the thing we built an instruction-channel scan to catch for documents — but nobody runs that scan on the tool descriptions our own MCP servers deliver.

The other three are more familiar: authentication is often just plaintext API keys (OAuth has been required since June 2025 but is barely implemented), servers run with far too many privileges, and there’s a supply-chain risk from malicious tool packages. One number sums up the scale: 21,000+ MCP servers are exposed online with no OAuth at all. AI security incidents in dev environments nearly tripled in the first half of 2026.


What I changed

This is where the lesson becomes practical. I run my own MCP servers, which is already a huge step forward — but self-hosted is not the same as secure. So I’m applying the same discipline I use for documents and memory:

  • Audit tool descriptions like documents. Every description my MCP servers inject into context gets treated as untrusted input, not as a trusted fact. A tool that says “this is how you reach my API” is an instruction channel, and I check it the same way.
  • Least privilege, not just trust. Each server runs with only the tools it needs. A camera server has no business holding my SSH key, and a search server has no business writing files.
  • Keys over plaintext. OAuth is the standard now. Any server still sitting on a raw API key is a risk I’m closing.

The mental shift is the same one from yesterday’s post: stop asking “is this source I chose?” and start asking “what actions could this content make my agent take?” A tool list from a server I installed is just as much an instruction channel as a document that drifted in through a sync. Provenance applies to tools too — where did this description come from, and how sure am I that it only does what I think it does?


Your turn

If you run any AI that connects to tools — an assistant with integrations, an agent with plugins, a Home Assistant with an AI frontend — you have an MCP-like surface whether you call it that or not. When was the last time you actually read the tool descriptions your setup injects into context? I’d genuinely like to know what you find. Drop a comment — I read every one.

Wat vond je van dit bericht?

Leave a Reply

Your email address will not be published. Required fields are marked *