Thirteen Model Context Protocol CVE records went into NVD from a single assigning authority inside a thirteen-second window on August 27; a fourteenth, scoring 10.0 against Argo CD’s MCP server, followed on August 29 — and the flaw it describes had a public GitHub advisory, and a fix, eighteen days earlier with no CVE attached.
What happened
On Saturday, August 29, 2026 at 2:16 p.m. UTC, NVD published CVE-2026-82456 against argoproj-labs’ argocd-mcp. VulnCheck is the assigning authority and supplied both scores: CVSS 4.0 base 10.0 and CVSS 3.1 base 10.0, the v3.1 vector AV:N/AC:L/PR:N/UI:N/S:C/C:H/I:H/A:H. The weakness is CWE-1327, binding to an unrestricted IP address. The record’s description, in the CNA’s own words: “argocd-mcp 0.8.0 binds its HTTP transport to every network interface and accepts MCP sessions without requiring caller credentials when ARGOCD_API_TOKEN is configured.” An attacker who reaches the port initializes a session with no credentials and then operates Argo CD with the operator’s stored token, creating applications and synchronizing attacker-supplied manifests into a Kubernetes cluster.
That record is two days old. What it led us to is four days old, and the dates matter, so here they are plainly. On Thursday, August 27, 2026, between 5:20:51 p.m. and 5:21:04 p.m. UTC, VulnCheck published thirteen CVE records against thirteen separate MCP projects: CVE-2026-81091 through CVE-2026-81102, twelve consecutive identifiers, plus CVE-2026-81735. The affected projects are mcp-use, mcp-go, Apify’s actors-mcp-server, mcp-router, pg-aiguide, ToolUniverse, rails-mcp-server, Telnyx’s telnyx-mcp, Timescale’s tiger-slack and tiger-gh-mcp-server, Airtable’s airtable-mcp-cli, Dropbox’s mcp-server-dash, and ByteDance’s UI-TARS-desktop.
Severity Daily published a page on CVE-2026-81735 on August 28, treating it as a standalone record. It was one of thirteen, published in the same thirteen seconds, by the same assigner. That context was not in our page and should have been. A dated note has been added there pointing here.
The weakness distribution across those thirteen is the most useful thing in the set. Five carry CWE-346, origin validation — mcp-go, pg-aiguide, tiger-slack, tiger-gh-mcp-server, and mcp-server-dash. Three carry CWE-306, missing authentication for a critical function — mcp-router, telnyx-mcp, and UI-TARS-desktop. Two carry CWE-918, server-side request forgery — mcp-use and actors-mcp-server. The remaining three are one each: CWE-94 code injection in ToolUniverse, CWE-78 OS command injection in rails-mcp-server, and CWE-200 information exposure in airtable-mcp-cli. CVSS 4.0 base scores run from 2.3 to 10.0.
The sweep is not confined to one assigner. An NVD keyword search for “MCP” limited to records published between August 25 and August 30, 2026 returns 31 results. GitHub’s security advisory program supplied a cluster on August 25 alone, including PraisonAI, mcp-shell, nextcloud-mcp-server, browse-mcp, genieacs-mcp, and the MCP PHP SDK. VulDB contributed records on August 27 and August 28. VulnCheck added mcp-fetch on August 26 and NousResearch’s hermes-agent, at CVSS 4.0 base 9.0, on August 28. That 31 is a floor rather than a count: keyword search matches description text, and two records in the August 27 batch — CVE-2026-81097 and CVE-2026-81101 — do not contain the string and do not appear in the result.
None of these records states that exploitation has been observed, none appears in the KEV catalog, and no federal deadline attaches.
Why it matters
Thirteen records in thirteen seconds is not thirteen discoveries. It is one query run against a class of software, and the shape of the result tells you what the query was. Someone enumerated MCP server implementations, checked what address each one binds by default and whether it validates the host header a request names, and filed what came back. The productivity of that exercise — thirteen hits across projects maintained by Dropbox, Airtable, Telnyx, Apify, Timescale, and ByteDance — is the finding. These are not obscure hobby repositories. They are connectors shipped by companies with security teams.
We made the structural argument about MCP’s transport shift in the UI-TARS page and will not repeat it. The new information is in the distribution, and specifically in the low-scoring records, which are where the useful lesson sits. Dropbox’s mcp-server-dash scored 2.3 on CVSS 4.0, the lowest in the batch. Read what it did: it bound its listener to the loopback address. That is the correct, widely recommended mitigation — the thing every hardening guide tells you to do, and the thing an engineer who had read the UI-TARS advisory would have done. It was still reachable, because per the CNA record it “never checked the host a request named,” so “a name that had been pointed at the loopback address still reached the listener while carrying the attacker’s host name.” A web page in a browser on that machine could drive the server and invoke its tools under the Dropbox credential the server holds.
That is DNS rebinding, and it is the reason five of the thirteen carry CWE-346 rather than CWE-306. The two defects are usually discussed as one problem — “the MCP server was exposed” — and they are not. Binding to loopback closes the network path and does nothing about the browser path, because the browser is already on the loopback interface. A defender who audits bind addresses across their MCP fleet, finds them all set to localhost, and marks the task done has completed half of an audit and will believe they completed all of it. The second half is whether the transport validates Host and Origin, which is not visible from a port scan, not visible from a process listing, and only visible by reading the transport configuration or the code.
Read the score spread as a warning rather than a ranking. Sorting an MCP inventory by CVSS and working down addresses the servers that did nothing and defers the ones that did something insufficient — which is close to backwards, because the projects that bound to loopback are the ones whose operators believe they are safe.
The argocd-mcp record carries a separate problem worth naming on its own. GitHub’s advisory for this flaw, GHSA-rp45-5x3v-48mr, was published on August 11, 2026, credits the reporter shmulc8, names 0.8.0 as affected and 0.9.0 as fixed, and carries the same CVSS 10.0. That advisory states “No known CVE.” The CVE record arrived eighteen days later, from a different assigner. NVD’s record for CVE-2026-82456 references the GitHub advisory; the GitHub advisory does not reference the CVE. For eighteen days a critical, fixed, publicly documented flaw in a Kubernetes deployment tool existed with no CVE identifier — which means it was visible to anyone reading GitHub advisories and invisible to every scanner, SBOM tool, and vulnerability feed that keys on CVE IDs. The fix shipped in 0.9.0 during that window. Organizations that patch from advisories got it; organizations that patch from CVE feeds did not know there was anything to get.
What to do
Build the inventory first, because most organizations do not have one. MCP servers are installed by developers, run as user processes, and rarely appear in a CMDB. Look for listening processes started from npm and pip packages with “mcp” in the name — on developer laptops, build agents, and any host running an internal agent framework.
For each, answer two separate questions. What address does it bind — loopback being necessary and not sufficient? And does the transport validate the Host and Origin headers? If that answer is no or unknown, the server is reachable from any web page the operator visits, whatever the bind address.
Where fixed versions are named in the records, they are specific. Upgrade argocd-mcp to 0.9.0 or later. Upgrade mcp-router to 0.6.3 or later, which defaults the host to loopback and refuses to start without a token when given a non-loopback host. For rails-mcp-server, 1.6.1 restricts what the sandbox permits and 2.0.0 removes the affected tool entirely. Look up the remaining identifiers at services.nvd.nist.gov by CVE ID; every record in this batch names its fixed release in the CNA’s description text.
Then treat any MCP server that holds a credential as a service with that credential’s privileges, and scope the token accordingly. The argocd-mcp record is a clean illustration: the flaw grants no Argo CD access of its own, it grants whatever ARGOCD_API_TOKEN was given. A read-scoped token would have produced a materially smaller incident than an admin-scoped one, and nothing about the vulnerability would have changed.
Sourcing note
Checked: the NVD records for CVE-2026-82456, CVE-2026-81091 through CVE-2026-81102, and CVE-2026-81735, each retrieved individually by CVE ID from the NVD 2.0 API, which supplied every timestamp, score, vector, weakness, and quoted description above; VulnCheck’s public advisory page for argocd-mcp; and GitHub advisory GHSA-rp45-5x3v-48mr, which supplied the August 11 publication date, the reporter credit, the affected and patched versions, and the “No known CVE” status. Counts of the wider sweep come from NVD keyword searches over the stated date range, run against the API.
Could not reach: api.github.com returned HTTP 403 to automated requests, so the advisory dates were read from the rendered advisory page rather than the API, and we could not retrieve the repository’s release timestamps to confirm when 0.9.0 actually shipped. cisa.gov returns HTTP 403 to automated fetching, so the absence of any KEV entry here was confirmed through NVD’s own CISA fields rather than the catalog page.
Unresolved: whether the August 27 batch reflects a single coordinated research project and, if so, whose. VulnCheck is the assigning authority for all thirteen, but the records do not uniformly name a discovering researcher, and we found no writeup describing the sweep as a whole. Also unresolved: why GitHub’s advisory carried no CVE for eighteen days, and whether the two records will be reconciled. No exploitation of any identifier in this story has been reported by any CNA, vendor, or agency as of publication.