Leroy H. Mason (Elior Malak) (@Lhmisme) header
Leroy H. Mason (Elior Malak) (@Lhmisme) profile picture

Independent AI security researcher. Built mcp-authscan, a focused auth scanner for live MCP servers. Finds the boring failures that actually ship.

Georgia, USA
Joined August 2009
Photo by Leroy H. Mason (Elior Malak) (@Lhmisme)Photo by Leroy H. Mason (Elior Malak) (@Lhmisme)Photo by Leroy H. Mason (Elior Malak) (@Lhmisme)Photo by Leroy H. Mason (Elior Malak) (@Lhmisme)Photo by Leroy H. Mason (Elior Malak) (@Lhmisme)Photo by Leroy H. Mason (Elior Malak) (@Lhmisme)Photo by Leroy H. Mason (Elior Malak) (@Lhmisme)Photo by Leroy H. Mason (Elior Malak) (@Lhmisme)Photo by Leroy H. Mason (Elior Malak) (@Lhmisme)Photo by Leroy H. Mason (Elior Malak) (@Lhmisme)
Two different checks get collapsed into one, and the difference decides where your credentials go. Metadata consistency: the authorization server metadata describes the server it was fetched from. Credential ownership: the issuer it names is the issuer your credentials were provisioned for. The first is enforced widely. The second is opt-in. GHSA-qx49-fqc8-xw99 and GHSA-vqqp-x5qm-5qp4 both land on the same residual, and the advisory says it plainly: upgrading changes nothing until you also pass issuer. FastMCP accepts issuer= from 4.1.2 and defaults it to None. The MCP SDK's own client provider does the same. No pin means no trust anchor, so a server advertises its own authorization server and a pre-provisioned client secret goes there. No new vulnerability here. The gap is a default. Read against source and demonstrated on a real pinned client. The bench is public and anyone can re-run it. mcpauth-scan.com
1
2
26
Today an mcp-authscan report on a real AI server became something a stranger can check end to end, from a public link, without asking me anything. The server is tollbooth-authority.fastmcp.…, scanned with the operator's written permission, publication included. Zero findings. Four checks were set aside, and each one has the reason written next to it, so a clean result is never a blank one. Put the report through the checker and you get three answers back: MATCH, SIGNED BY 9b79936f, and MANIFEST MATCHES PIN. Nothing in the results was changed, the key that signed it is the one published on the site, and the report came from the released version of the tool. I produced it two ways, once from my working copy and once from a clean install of the published package. Different builds, same report hash. Cisco's scanner is built for the buyer checking its vendors, and that's a real job. mcp-authscan is built for the other side of the table: the founder whose deal is stuck in a security review. Its report is made to be checked by whoever they hand it to. Cisco and Ramparts publish what they cover. mcp-authscan also publishes an answer key, so anyone can see whether the grader gets the right answer when the answer is already known. It doesn't read code or tool content, and a matching hash tells you the result didn't change, not that the right check was run. A security claim should never be stronger than the evidence behind it. mcpauth-scan.com/verify Receipts over promises.
2
1
44
The report is here: raw.githubusercontent.com/LH… Drop it into mcpauth-scan.com/verify with my key, published at mcpauth-scan.com/keys.json, and you get MATCH, SIGNED BY 9b79936f, MANIFEST MATCHES PIN. Don't take my word for it.
1
17
A security claim should never be stronger than the evidence behind it. MCP AuthScan v1.1.32 is officially launched. Its policy gate returns CLEAN, CONDITIONAL or REJECT. REJECT exits non-zero, allowing a deployment pipeline to stop when an assessment fails its configured policy. But a gate is only as valuable as the evidence behind its decision. MCP AuthScan assesses live MCP deployments and their OAuth infrastructure, distinguishing what was demonstrated, what failed, what could not be observed, and what could not be tested. A timeout is not a passing control. Every report carries a cryptographic hash over its graded content. Reports support digital signatures and independent verification. A second party can recompute the report hash without trusting the person who produced it. The probe set and severity calibration are versioned separately from the tool, preserving the meaning of historical assessments. The software itself is held to the same standard. This release records 713 passing tests, zero failures, and identical SHA-256 hashes between the locally built wheel and the wheel downloaded from PyPI. No claim of perfect security. No manufactured certainty. Just measurable observations, transparent limitations and verifiable evidence. A security product should hold itself to the standard it asks of others. Open source. MIT licensed. Authorized testing only. pip install mcp-authscan==1.1.32 mcpauth-scan.com github.com/LHMisme420/mcp-au… #MCPSecurity #AgenticAI #DevSecOps #OAuth #ZeroTrust
1
1
58
If your AI server passed a security check last month, how would your buyer know it still passes today? Until now they'd have to ask the person who ran it, and take their word for it. mcp-authscan got two upgrades today. The evidence is signed now, along with the grades. You could already check that the results in a report weren't touched. Now you can check the raw observations behind them too. Change one detail and the public checker says SIGNATURE INVALID. Strip the evidence out and it says UNREADABLE. Every report also carries the previous run inside it. To see what changed between two scans, a buyer needs that one report, one fresh scan, and my public key. They don't need files from me, access to my server, or any trust in anything I run. The key that signs is published on the site. The checker is live and free, and it has already been used on a real report. So the next time someone hands you a security report, ask them what it said last time, and where you go to check that too. mcpauth-scan.com/verify Receipts over promises.
1
1
79
Every security report you've ever been handed is a PDF, and a PDF is just a claim from someone who wants you to believe it. The person who paid for it has no way to check a single line. As of today, anyone can check an mcp-authscan report in about ten seconds, for free, no account. Drop it in and you get two answers: whether anything in the results was changed, and whose key signed them. The key is public. You don't have to trust me. It won't tell you a server is secure. It tells you the report is the one I signed, and that if I got it wrong, you can catch me. Next time someone hands you a security report, ask them where you go to check it. mcpauth-scan.com/verify Receipts over promises.
2
155
Every MCP scanner I know of tests the same thing: can an anonymous caller get in. None of them ask the next question. Once a caller is authenticated, does the server check which service the key was issued for? Built that check into mcp-authscan today. It mints a token for a real sibling service on the same issuer, presents it to the target, and grades the answer. FAIL HIGH: the target accepted a token cut for another service. PASS: it refused. Both answers above, one command, same probe set and stamp. The limit, said plainly: it needs a sibling service to mint against. Without one the probe excludes and names the reason. I would rather it report nothing than guess. Locked is not the same as checked. mcpauth-scan.com
1
47
Day 30 of 30 Days of @boardyai, the last one. Today the sample report went live on mcpauth-scan.com. At the bottom of the page there's a link, "See a sample report." It opens a real report from a real scan of a real company's AI server, run with their permission. The company is AgentPayStore. Thank you to Shawn Lippert, who runs it, for clearing it to be shown with the name on it. The scan ran on 16 November. It came back with one finding, and fifteen checks were set aside, each with the reason written next to it. The report carries a fingerprint, so anyone can run the same checks on the same server and see if they get a match. For a month I've been saying you shouldn't have to take my word for a report. Now there's one you can open and check. @boardyai was in this with me the whole way, on the grader, the answer key and the site. mcpauth-scan.com Receipts over promises.
1
40
Last month an attacker got into 56 customer projects at Mittwald, a German web hosting company, and they never broke the hosting console. They went through the AI connector standing in front of it. Mittwald had built that connector so customers could plug an AI assistant into their hosting account. The keys to the console passed through it. The attacker was inside for eight days. Mittwald's director said the logs show seconds between steps that would take a person far longer. Mittwald hasn't said more than that about how it happened, and I'm not going to guess. What I keep thinking about is where they got in. A lot of companies shipped one of these connectors this year, fast, so their customers could use AI with their product. If you're one of them, that connector is a way into your product, and attackers are finding weak spots faster than most teams can patch them. This is the time to have it checked, before someone else checks it for you. That's what I do now. I look at how your AI connector handles sign-in, what it shows to someone who hasn't signed in, and what changes after you ship a fix. The report carries a fingerprint anyone can recheck by running the same checks on the same server, so you can hand it to a customer or a buyer and they don't have to take your word for it, or mine. One server is $1,750. A full assessment with one retest is $3,500. We agree the scope in writing before anything runs, the results come to you first, and you get time to fix things before anything is said publicly. You can send a scope request straight from the site now, no call needed. mcpauth-scan.com Receipts over promises.
50
Attackers are now reading security fixes the day they ship and building attacks from the change, with no human involved. That's from the latest report by the Google Threat Intelligence Group. Every one of those fixes changes what your server shows the world. The moment it ships, whatever security report you had from last month is describing a server that no longer exists. Google's answer is to use Gemini and other frontier models to find and fix flaws before attackers get to them. That's good work and it's needed. What I keep thinking about is the picture you had before the fix. mcp-authscan keeps that picture. Every scan records a fingerprint of how your AI server handles sign-in. Scan, patch, scan again, and it names exactly what moved. A fix that closes one door can open another. You only find out if you kept the before picture with a hash on it. The same goes for the report you hand a buyer. Someone else runs the same checks on the same server and gets MATCH, or a list of exactly what changed, or NOT_REPRODUCIBLE. When attacks are automated and fixes land every week, a report nobody can recheck is just a claim from last month. What it doesn't do: it doesn't read your code, doesn't read the patch, and doesn't find bugs in fixes. It's a point-in-time check on one server's sign-in setup. The part that matters here is the before and after. After every patch, ask whether your server changed, and whether it changed the way you think it did. mcpauth-scan.com Receipts over promises.
4
92
Every AI server security scanner I know of is built for the buyer. Cisco's is a good example. An enterprise points it at its vendors and gets a report. That's their side of the table, and it's a real job. mcp-authscan sits on the other side. It's for the company selling the AI server, the founder whose enterprise deal is stuck in the buyer's security review, who already ran a free scanner and got told that results you produced yourself aren't proof. So the report is built to be checked by the person you hand it to. It carries a hash over everything that was graded, with timestamps left out, so the same results always produce the same hash. Your buyer runs the same checks on the same server and gets a match, or gets a list of exactly what changed. They don't have to trust you, and they don't have to trust me. The grader can be checked too. It has an answer key I built with @boardyai: small test servers whose grades were written down before the scanner ever touched them. One command prints HIT or MISS next to each one. Right now it reads 5 hit, 0 miss. Anyone can clone it and run it, or catch me the day it gets one wrong. What that doesn't do: a matching hash tells you the result didn't change. It doesn't tell you the check was the right one to run. That part is on the person who signed the report. A scan you pay for once ends up filed away. A report the buyer can re-run is something you can put in front of every buyer who asks. mcp-authscan is MIT licensed and free to run. Receipts over promises. mcpauth-scan.com
2
1
74
Day 29 of 30 Days of @boardyai. I put three AI server scanners on one machine and pointed them at the same server. They don't answer the same question, and they're not really competing. mcp-authscan, mine, asks one thing: does this live server let a stranger in without a login? Against a test server I stood up with no login at all, it came back HIGH. That's the picture. Against hosted.flowstatemcp.com, scanned with the operator's written permission, it came back PASS with 10 checks passed. Cisco's scanner looks inside. It reads the tools, prompts and resources a server offers, including tool descriptions written to trick an AI. It opens with a newer handshake than my little test server speaks, so I didn't get a content scan out of it today. That says something about my test server, nothing about Cisco. mcp-scan looks at the AI setups installed on your own computer. It doesn't take a web address, so there was nothing to point it at. So it's three lanes. Is the door locked. What's behind it. What's set up on your laptop. Owning one of these doesn't answer the other two, and anyone buying a scanner should know which question they're paying for. mcpauth-scan.com Receipts over promises.
7
54
Your MCP server may have more than one front door. Every scanner I know tests the endpoint you point it at. If your host also answers /sse next to /mcp, nobody looks there. mcp-authscan does now. One request, no credential. 401 or 403 means that transport enforces auth. A 200 that streams, or carries a JSON-RPC reply, is a finding under MCP-TRANSPORT-1. Anything I cannot read is excluded, never passed. The screenshot is that check against a test host I stood up that answers with no auth. It shows the check working, not a finding against anyone's server. A missing auth challenge on the front door nobody tested is not a smaller problem than one on the front door everyone tests. Run it on your own host: pip install mcp-authscan. mcpauth-scan.com
2
2
79
Your AI app can pass every sign-in check and still hand its login to a stranger. That's the run in the picture. A decoy sign-in service sits at a different address than the AI server. The server tells the app to go there. The app goes and sends its login. The last line says captured: present, which means the login arrived at the decoy. This isn't a new discovery. fastmcp says it at the top of the run, in its own warning: if the app isn't told which sign-in service to trust, the server decides where the app's login goes. They've said that setting will be required in 3.0, and I'm glad they wrote it down so plainly. What the run adds is the receipt. Same app, same test, one change. Left on the default, it failed in every release I've run it against. When the app was told which sign-in service to trust, it passed. A check that the sign-in details agree with themselves doesn't tell you where your login actually went. Those details can agree perfectly while pointing at a place nobody chose. If you build with fastmcp, tell your app which sign-in service to trust now rather than waiting for 3.0. Anyone can repeat this with one install and one command. mcpauth-scan.com Receipts over promises.
1
1
3
80
The AI server security scores going around are measured hardest on the servers that needed testing least. A properly locked server answers a stranger with one word: no. A scanner knocking without a login gets that no and has nothing to grade. An honest tool says nothing. And nothing gets read as clean. So the open servers get graded thoroughly. The locked ones, the servers companies actually put their data behind, come back blank. Today I found one of my own checks in mcp-authscan doing exactly that. It gave up the moment a server said "you need to log in," so a properly locked server could never finish a clean scan. It sat marked incomplete, every time. I fixed it the same day, with @boardyai helping me think it through. Now when a server asks for a login, that check is set aside as EXCLUDED, with what it would need written next to it: a response the server gives without a login. A locked server can finish a clean scan, and the report names every check it couldn't reach and why. A false all clear is worse than no scan. When a report says your AI servers are fine, the first thing to look for is how many checks actually ran. mcpauth-scan.com Receipts over promises.
2
36
Day 28 of 30 Days of @boardyai. Every scanner in this space tells you what it looks at. Cisco publishes its coverage. Ramparts publishes its coverage. Nobody publishes whether their grader gets the right answer when the answer is already known. So today I built an answer key for mine, with @boardyai helping me work through it. It's four small fake AI servers, each with its grade written down before the scanner ever touches it: A server with no login at all. Expected MEDIUM. A server that refuses everything. Expected INCONCLUSIVE. A properly closed server that publishes its sign-in details. Expected PASS. A plain web page that isn't an AI server at all. Expected INCONCLUSIVE. One command starts each one, scans it, prints HIT or MISS next to it, and gives a score. Today it reads: ANSWER KEY 4 hit 0 miss You don't have to take my word for that. A vendor or a reviewer can run the key themselves and get the same score, or catch the grader the day it gets one wrong. It works the same way as the report hash. The tool can be caught. mcpauth-scan.com Receipts over promises.
1
50