跳到正文
DEV Community· Cor E·· 15 小時前AI 評分53

62% MCP 伺服器向未經身份驗證的用戶端公開工具清單,作者提醒這不等於已遭入侵

62% of MCP Servers Will Show Their Tools to Anyone. That's Not the Scary Part.

AI 導讀

Spellwarden 探測官方 MCP Registry 背後的 30,402 個主機;在有回應的伺服器中,62% 向未經身份驗證的用戶端公開工具清單,其中數千個伺服器公開非唯讀或標示為破壞性的工具。

正文

Someone finally counted, and the number is what you'd expect from a protocol that shipped before its threat model did.

Spellwarden probed 30,402 hosts behind the official MCP Registry. Of the servers that responded, 62% handed over their tool list to unauthenticated clients. Thousands of those expose tools that aren't read-only, or that are declared destructive. That's the headline. It got exactly one point on Hacker News and zero comments, which tells you something about how the industry digests this kind of finding: it nods and scrolls.

Context

Nothing about this is a new class of problem. Unauthenticated enumeration of a service's capabilities is the same story we told about open Redis instances, exposed Elasticsearch clusters, and Docker daemons listening on 2375. A new protocol arrives, people stand up servers fast, auth is "something we'll add," and eventually someone runs a census and publishes a percentage.

What's different is the consumer on the other end. A leaked tool list on a normal API is reconnaissance for a human attacker. A tool list on an MCP server is meant to be read by an agent, and agents act on what they read. That changes the stakes, at least in theory.

Hype check

Here's where I'd slow down. The report is a census of exposure, not an observed attack. No LLM was actually compromised. Anyone writing "MCP servers are being hijacked in the wild" off this data is making it up.

And look at the finding that cuts against the scary narrative: instruction-like text in tool descriptions turned up in roughly 1 in 100 open tool lists. Tool-description poisoning is the attack everyone loves to demo at conferences, and in this dataset it's uncommon. That's the part the breathless takes will skip, because "1% of descriptions look odd" doesn't sell a webinar.

But I'd also push back on the opposite reflex, the "so it's fine" shrug. Listing tools without authentication isn't the same as invoking them without authentication, and the summary doesn't say how many servers would actually execute a destructive call for an anonymous client. That's the number I want, and it's the one this kind of census can't easily give you without crossing from probing into poking. Exposed tool list plus declared-destructive tools is a bad smell. It isn't a proven breach path.

Who benefits from the loud version of this story? Anyone selling agent-security tooling. The framing "huge attack surface for agents connecting to untrusted MCP servers" is true as far as it goes, and it's also exactly the sentence that opens budget conversations. Read the data, not the implication.

Implications

The practical takeaway for developers is boring and correct. If you run an MCP server, authenticate it. Default to read-only. Mark destructive tools as destructive and then actually gate them, because a label nobody enforces is documentation, not a control.

The takeaway for anyone building agents is the one I'd stress harder, because the flip side of "servers are open" is "your agent will happily connect to whatever is listed." The attack surface in that summary is described from the agent's side: connecting to untrusted servers. A registry is a trust signal people lean on without examining. Being in the official registry tells you a server exists. It doesn't tell you it's safe, and this census is a decent reminder of that gap.

For security teams, treat MCP endpoints like any other internet-facing service in your asset inventory. I'd bet most orgs couldn't tell you which agent talks to which MCP server today. That's the real exposure, and no census captures it.

I'll admit my first instinct on reading 62% was to assume it was inflated by a pile of toy servers and abandoned demos. The summary doesn't break that down, so I can't tell you whether that instinct is right. A registry full of hobby projects and a registry full of production integrations are very different risk pictures, and both would produce the same headline number.

Open question

If most of these servers are unauthenticated by neglect rather than design, who owns fixing that: the server authors, the registry that lists them, or the agent frameworks that connect to them without asking questions?

— Cor, Skyblue Soft

Sources


AI-assisted draft or imaging, human-curated, reviewed and edited.

來源:DEV Community · dev.to