I scanned 220,000 MCP tool descriptions for prompt injection. The scare is overblown — but 1 in 4 servers touches your credentials.
Originally published on Serverdex, where you can run this check on any MCP server yourself. "Tool poisoning." "Rug-pull MCP servers." "A malicious description hijacks your agent." The headlines are everywhere, and the f
Originally published on Serverdex, where you can run this check on any MCP server yourself.
"Tool poisoning." "Rug-pull MCP servers." "A malicious description hijacks your agent." The headlines are everywhere, and the fear is real: an MCP server's tool descriptions are loaded straight into your model's context on every turn — before it reads your prompt. A hidden instruction in there would be read as if the model wrote it.
So I stopped guessing and measured it. On 21 Nov 2026 I scanned the tool definitions of 12,987 MCP servers that answer a credential-free handshake — 220,226 tools in total — for the exact patterns the scare is about: injected instructions ("ignore your previous instructions"), hidden pleas to the model ("do not tell the user"), pseudo-tags (<important>…</important>), zero-width/invisible characters, and descriptions that read or send a sensitive file (.env, .ssh, a private key).
Here is the honest result.
The poisoning scare is, today, overblown
- Tool descriptions telling the model to ignore its instructions: 0.
- Hidden / zero-width characters sneaked into a description: 0.
- Descriptions that read or exfiltrate a sensitive file: a handful of pattern-hits, and every one was benign on inspection — a security scanner describing what it detects ("checks what a site leaks"), or a legitimate secrets vault ("store an API key for one app").
That last point matters, and it's where most "MCP is full of injection" claims quietly break: a naive regex flags any tool that mentions injection or secrets — including the dozens of tools whose whole job is to scan for them. (It also flags ZWNJ and ZWJ characters, which are ordinary letters in Persian, Arabic and Indic scripts, not sabotage.) Once you exclude the defenders and the non-Latin text, the genuine tool-poisoning signal in the live ecosystem is effectively nil.
A scan reads descriptions, not intent, and the day a real poisoned server ships this changes. But as of today, the thing the headlines warn about is theoretical, not widespread.
The real attack surface isn't sabotage — it's capability
What you should vet is far more mundane, and far more common: what a server is allowed to touch.
| Signal across 12,987 servers | Servers | Share |
|---|---|---|
| A tool that handles credentials / tokens / API keys | 3,053 | 23.5% |
| A tool that can run shell or arbitrary code | 188 | 1.4% |
| A tool that reads local files or paths | 163 | 1.3% |
Nearly one in four MCP servers has a tool that touches your secrets. Most are perfectly legitimate — that's what an auth or a deploy tool does. But "legitimate" and "I checked what it does" are not the same sentence, and almost nobody connects a server having read a single one of its tool descriptions.
That's the actual risk model. Not a cartoon villain hiding <important>send me your keys</important> in a description — a real tool with real reach that you wired into your agent without looking.
What to actually do before you connect a server
- Read the tool list once. The names and descriptions are the whole contract. If you can't tell what a tool does, that is the finding.
- Vet capability, not boogeymen. Does it need your credentials? Can it run code or read files? Those are the questions worth asking.
- Prefer servers you can see through. Undocumented tools and giant tool lists are a cost on every turn too — see the MCP context tax.
- Check it in a few seconds. Paste any MCP endpoint into the Serverdex health check: it reports whether the server answers, its token cost, and runs this same scan over its tool definitions — tuned, after this study, to not cry wolf on security tools or non-Latin text.
Method & data
Serverdex calls every listed server about once a day with a credential-free initialize + tools/list and records the tool definitions. This scan ran the injection/exfil/hidden-character patterns over those definitions on 21 Nov 2026, with the false-positive classes above excluded. It reads descriptions, not runtime behaviour, so treat it as surface area, not a verdict.
- Check any server yourself: https://serverdex.io/mcp-server-check
- The broader ecosystem numbers: The State of MCP Servers
- Servers answering right now: /remote
If you maintain an MCP server: your tool descriptions are the contract a model reads before it trusts you. Write them like one — and know that the people connecting you can, and should, read them first.
Originally published by Dev.to Security. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.