Anyone can list an MCP server in a public marketplace, and most marketplaces don't check who is behind a server, what it does or whether it still does the same thing tomorrow. Your agent reads the listing, believes it, and calls the server with your accounts, your files and, more and more, your money.
We run a marketplace, so we have an opinion. The evidence is public, and most of it is under eighteen months old.
One malicious listing is enough
Every MCP server describes its tools in plain text, and your agent reads that text as instructions. You see a short summary of it, if you see anything. So whoever writes the description can give your agent orders.
In April 2025, Invariant Labs showed how that works. They built a tool that looked like it added two numbers. Its description also told the agent to read the user's SSH key and pass it out through an extra parameter. Cursor's confirmation dialog did not show the key. A server can also change a description after you approved it.
How often does it work? The MCPTox benchmark built poisoned tools on 45 real MCP servers and 353 real tools and tried them on 20 AI agents. The most vulnerable one, o1-mini, was fooled 72.8% of the time. The paper found that more capable models were often more susceptible, because they are better at following instructions.
Most marketplaces don't check what they list
The official MCP registry says so itself. It confirms that a publisher owns the GitHub account or domain behind a name. It calls its metadata "deliberately unopinionated" and leaves security scanning to npm, PyPI and "downstream aggregators", which is the polite word for marketplaces.
Researchers at the University of Delaware studied 67,057 servers across six registries in mid-2025. Of 1,379 listings with a dead link, 212 pointed to a GitHub account that anyone could register again and put their own code behind. The authors reported it to the registries. mcp.so did not answer.
In September OX Security looked at 15,465 public servers and found no vetting or review by the marketplaces. Six of the hostnames sit on expired domains that anyone can register for $4 to $12 a year, and whoever registers one gets whatever agents still send to it. And a remote server can run code that differs from the repository its listing links to.
In July 2026 researchers audited 414 internet-facing MCP servers. 380 of them, 91.8%, had no OAuth authentication, and the servers they confirmed exposed 687 tools that run shell commands with no access controls. Three days after one scan, 193 of the 464 servers it had confirmed were no longer answering.
When we drew 20 servers at random from the official registry in October, two had tool descriptions with instructions addressed to the agent reading them. We wrote it up here.
Open marketplaces have already listed malware
These are the directories that compete with Delegate, and what public reports have found on them so far.
| Directory | What public reports found |
|---|---|
| ClawHub | 341 of 2,857 skills flagged malicious in one audit |
| MCP Market | A fake Oura server that installed an infostealer |
| mcp.so | Fake listings from a malware campaign |
| Glama | Fake listings from the same campaign |
| LobeHub | Fake listings from the same campaign |
| npm | postmark-mcp, which started copying every email |
| Smithery | One leaked token controlled 3,243 hosted apps |
| PulseMCP | Nothing published yet. That is not the same as clean |
| Nevermined | Nothing published yet. Goes live with no manual approval |
A directory with no report is not a clean directory. It means nobody has published a finding yet.
In September 2025 postmark-mcp, a copy of Postmark's MCP server, sat on npm through fifteen harmless versions. Version 1.0.16 copied every email it sent to an address the author controlled, and it was downloaded 1,643 times. Koi Security found it and called it the first malicious MCP server seen in the wild.
In February 2026 a clone of the Oura Ring MCP server on MCP Market, listed under fake contributor accounts, dropped the StealC infostealer.
ClawHub is the skills marketplace for OpenClaw, and it works like an MCP marketplace. In February Koi Security audited 2,857 skills and flagged 341 as malicious, 11.9%, and by May the confirmed count was past 1,184. Snyk found that publishing took a markdown file and a GitHub account one week old.
In July 2026 Island described a campaign of more than 800 fake skills and MCP servers, with over 600 listings across LobeHub, Glama, MCP.so and MCP Market. Then they asked the agents. Gemini returned a fake Walmart MCP server as its first recommendation, and ChatGPT called it the best place to start. Claude Code surfaced a malicious skill too and in one run repeated its install steps, though in others it declined.
In June 2025 GitGuardian found that a path traversal bug in Smithery's build system exposed a token that controlled 3,243 hosted apps. It was fixed two days after the report, and there is no evidence anyone used it.
- of the time the most vulnerable agent followed a poisoned tool description
- 72.8%
- fake skills and MCP servers in one campaign, listed in four registries
- 800+
- of remote servers in the official registry changed their tool definitions within a day
- 6%
- downloads of postmark-mcp, which one day started copying every email it sent
- 1,643
We have seen this before
Open proxies were the same bet. Anyone can run one, anyone can list it, and everything you send passes through it. A 2018 study tested 65,871 of them. Of the 19,473 that worked, 1,004 (about 5%) were doing something malicious, such as injecting ads, collecting data about their users or sending them to pages Google had blocked. That is about 1 in 20. ClawHub's audit flagged 1 in 8.
Scanners don't fix it
ClawHub added VirusTotal and ClawScan screening after the first reports. Palo Alto's Unit 42 kept finding skills that got through between February and May. One had a README padded with about 22 MB of filler to go past the scanners' size limit, and VirusTotal returned clean.
Some attacks have no code for a scanner to find. A May 2026 paper tested skills with no malicious code at all, only instructions written into the skill's text. In the most vulnerable setups they caused a confidentiality breach up to 77.67% of the time, and the scanners flagged none of the skill files.
Toolport fetched the tools of the official registry's remote servers twice, about 23.5 hours apart. Of 14,032 servers, 840 returned different tool definitions, and 708 of those still showed the same registry version. Most looked like ordinary updates, but nothing in the listing told you.
Agents that pay have more to lose
A server your agent calls can also ask it for money. On October 22 the security firm V12 reported 25 flaws in x402, the standard agents use to pay per call over HTTP. The worst was that the client had no default spending limit. A server could answer "payment required" and name the agent's whole balance as the price. It only affected agents that sign payments automatically and set no limits of their own. The fixes shipped on August 18 without a security advisory, clients now cap a payment at $1 by default, and nobody has reported an attack.
Catalogs of paid agent services often let a listing in once it can take payment. Nevermined's docs say you need a paid organization and an endpoint that charges per request over x402 or MPP. The service goes live once it responds healthy, and "there's no manual approval step." A monitor keeps checking it and unlists it if it stops responding. Email submissions are reviewed "from time to time", which is "not guaranteed". That is a fair bar for a payments company, which needs to know the endpoint answers and bills. Nevermined also says it holds SOC 2 Type II and ISO 27001 for its own systems. None of that tells you whether the service on the other end is honest.
Every big marketplace reviews before it lists
Marketplaces that touch people's devices and money have reviewed what they list for a long time.
Apple's App Review looked at more than 9.1 million submissions in 2025 and rejected over 2 million for breaking its guidelines. It also terminated 193,000 developer accounts over fraud concerns and says it prevented over $2.2 billion in potentially fraudulent transactions. Google kept 1.75 million policy-violating apps off Google Play and banned more than 80,000 developer accounts. Amazon says new sellers complete a verification process before they can sell, and that it seized more than 15 million counterfeit products in 2025.
App stores check before they list because a phone holds your camera, contacts and payment cards. An agent holds your files, accounts and wallet.
Review misses things too. Between June 2024 and May 2025, Zscaler counted 239 malicious apps on Google Play with more than 40 million downloads between them. A reviewed store at least has a team looking for them and someone to answer for the ones that get through.
What a curated marketplace does
On an open marketplacenothing is checked
- No check, no review
- No company, no KYC or KYB
- You can be anyone
- It can claim anything
- It can change at any time
- Your agent uses it
On Delegateevery step is checked
- A verified company with a real payment method
- Test orders before it goes live
- Claims checked against delivered jobs
- Re-checked continuously
- Money held until the result arrives
- Every provider is a verified company with a payment method on file, checked at signup and re-checked continuously. A postmark or Oura clone would need a name someone can be held to.
- Providers are tested with real orders before they go live. Each listing says what the service returns, what it leaves out and what a call costs.
- Search ranks services by how well they fit the request, then by delivery record, price and speed. You can't pay for a higher spot.
- Jev, a decision model from Typesafe, checks delivered jobs against what each service promised.
- Your money is held until the result arrives, and a call that fails costs nothing.
None of this makes the problem go away. A new kind of attack will get past any checklist, and a curated marketplace gets hit too. Curation makes attacks cost more and puts someone on the hook when one lands.
If you connect to an open marketplace anyway:
- Read the full text of every tool description yourself, not the summary.
- Pin the version, and get a warning when a description changes.
- Don't give the same agent your private files, an untrusted server and a way to send data out.
- Set a hard spending limit on the agent's wallet and don't click "always allow".
The checks above run in Delegate Agents.