Two MCP Bugs and a Five-Year-Old ISPConfig Flaw Got Exploitation Flags Today: The MCP Python SDK Session Bug, MCPHub's Non-Admin RCE and ISPConfig SQL Injection. Here Is the Fix
What to do first. Three bugs entered VulnCheck's catalog of exploited vulnerabilities today, November 20, 2026, each flagged as seen by VulnCheck's own canaries. Two of them are in software that runs AI agent tooling, built on the Model Context Protocol (MCP). The third is a hosting control panel bug from January 2021. None is in CISA's KEV catalog. If you run any of these, act today.
MCP Python SDK (mcp on PyPI) below 1.27.2: upgrade. Since the same package has had other fixes after 1.27.2 (see below), move to a current release: at least 1.30.0 on the 1.x line, or 2.2.0 or later.
MCPHub (@samanhappy/mcphub) 0.12.14 and earlier: upgrade to 0.12.15 or later (the current npm release is 1.2.6). Until you do, make sure every account on the hub is an admin you trust, close self-registration, and keep the hub off the internet.
ISPConfig below 3.2.2: upgrade to 3.2.2 or a later release. On 3.1 (3.1.12 and up), ISPConfig published a patch you apply with its patch tool as root (patch name 3115_sqlinjection; 321_sqlinjection for 3.2). Then log in once as admin, as ISPConfig's notice asks.
What each bug does
MCP Python SDK, CVE-2026-52869 (GHSA-jpw9-pfvf-9f58, CVSS 7.1). This is the official Python SDK for building MCP servers and clients, pulled about 331 million times in the last month, per PyPI figures our package checker reads. The bug is on the server side. When a server built on the SDK used the SSE transport or the stateful Streamable HTTP transport, it routed each request to an existing session by the session ID alone (the session_id query parameter, or the Mcp-Session-Id header) and never checked that the request came from the same authenticated client that opened the session. Per the advisory, any client holding a different valid bearer token and a known session ID could inject JSON-RPC messages into someone else's session, and on Streamable HTTP it could also read the replies.
The advisory is precise about who is exposed: only servers that use one of those HTTP transports and the SDK's bearer-token authentication. Servers on stdio, stateless Streamable HTTP, or no authentication are not affected. Session IDs are random UUIDs, so an attacker needs one from somewhere else, such as logs or traffic it can observe. The SSE transport was affected since the first release, Streamable HTTP since 1.8.0. The fix in 1.27.2 records which client (and, where the token verifier supplies them, which issuer and subject) opened each session and answers anyone else with a 404. If many of your users share one OAuth client, as with a hosted MCP client or a gateway, the advisory says to make your token verifier fill in AccessToken.subject from the token's sub claim, or sessions are only separated per client, not per user.
This is not the bug we wrote about in September. Our post A Malicious MCP Server Could Name Its Own Login Service, and the Official Python SDK Believed It covered GHSA-qx49-fqc8-xw99 (now CVE-2026-115563), a client-side OAuth flaw fixed in 1.30.0 and 2.2.0. Same package, different code, different fix. That is why we say to upgrade to 1.30.0 or 2.2.0 or later rather than stopping at 1.27.2: 1.27.2 closes today's exploited bug and leaves the September one open.
MCPHub, CVE-2026-79748 (GHSA-mx89-jjx9-gjr8, CVSS 9.9). MCPHub is a self-hosted hub that runs many MCP servers behind one endpoint (about 2,700 GitHub stars, about 9,200 npm downloads last month). Adding an MCP server to it means giving it a command to run. Per the advisory, the endpoints that add or edit a server, POST /api/servers and PUT /api/servers/name, check that you are logged in but not that you are an admin, and do not restrict the command. So any logged-in non-admin account could register a "server" whose command is /bin/sh with arbitrary arguments, and MCPHub ran it at once, as the user MCPHub runs as. The advisory says that is commonly root in the published Docker image and in npx and systemd setups. The edit endpoint also had no ownership check, so a non-admin could rewrite a server the admin had set up. Credited to GitHub user offset; fixed in 0.12.15.
Who is exposed: any MCPHub on 0.12.14 or earlier with more than one account, or with registration open. The hub holds API keys for every upstream MCP server it fronts, in mcp_settings.json, so a takeover leaks those too. A single-admin hub with no other accounts is not reachable through this bug, but that is a thin wall: one leaked low-privilege login is enough.
ISPConfig, CVE-2021-3021 (CVSS 9.8). ISPConfig is an open-source control panel that hosting providers and small businesses use to manage web, mail and DNS servers. NVD's record is one line: "ISPConfig before 3.2.2 allows SQL injection." ISPConfig's release notice of January 4, 2021 is equally brief, does not say which module or whether a login is needed, and credits Cyku Hong of DEVCORE. We could not find the technical details behind the flag, so we are not going to guess at the request. What we can say: a control panel's database holds the accounts and settings for every site and mailbox it manages, and a panel this old that is still unpatched is likely missing many other fixes too.
How sure is the exploitation?
One source for all three. VulnCheck's canaries flagged them, with VulnCheck's date_added of November 20, 2026. VulnCheck uses date_added for its earliest exploitation evidence, not for when an entry was typed in. The sighting details sit behind a paid VulnCheck subscription we do not have, so we cannot tell you what the requests looked like, how many there were, or where they came from. None of the three is in CISA's KEV catalog (checked November 20, 2026, against the catalog released November 18). VulnCheck marks ransomware use as unknown for all three.
Read the MCP Python SDK flag with care: that bug needs a valid token and someone else's session ID, which is hard to get from a canary. We do not know what VulnCheck's canary saw. Patch anyway; the upgrade is cheap.
What our own sensors can and cannot say
We checked our three edge instruments from October 21 to today: edge blocks (601,513 records), edge trap hits (100,586) and the origin honeypot (691).
MCPHub: zero requests to /api/servers or anything containing "mcphub". We run no MCPHub and no trap for it, so that zero means our sites were not targeted, not that nobody is scanning for hubs.
MCP Python SDK: zero requests to /sse or /messages. That is weak evidence at best. The bug is in the session ID and token checks, and our edge records keep the path without the query string or headers, so an exploit attempt against a real server would not look different from normal MCP traffic in our data.
ISPConfig: zero requests containing "ispconfig". ISPConfig usually listens on its own port (8080 by default), which our sites do not expose, and no exploit path has been published, so this is a blind spot, not a measurement.
None of our own pages matched these patterns, so there was nothing to remove from the counts.
What our checkers say
Jeevesus check-package on PyPI mcp version 1.27.1 returns verdict review and lists GHSA-jpw9-pfvf-9f58 (CVE-2026-52869) among the version's published vulnerabilities. On 1.27.2 that entry is gone and the September OAuth advisory is still listed, which matches the advice above.
Jeevesus check-package on npm @samanhappy/mcphub 0.12.14 also returns review, but for a different MCPHub advisory (GHSA-wf8q-wvv8-p8jf). It does not list CVE-2026-79748: the public vulnerability database it reads (OSV) has today's CVE without the npm package attached. Do not rely on our checker for this one; check your version by hand.
Dredd, our MCP server judge, returns review for samanhappy/mcphub, meaning we hold no record of it. Dredd judges MCP servers listed in the MCP Registry; MCPHub is a hub you host, not a registry server, and the MCP Python SDK is a library. Dredd does not flag either bug.
Am I affected? A self-check
MCP Python SDK: run pip show mcp (or check your lock file) in every service that hosts an MCP server. If the version is below 1.27.2, then look at how the server starts: if it uses SseServerTransport or StreamableHTTPSessionManager in stateful mode, together with the SDK's bearer-token auth, it is affected. Also check whether session IDs show up in your access logs or in URLs that get shared: for SSE, the session_id sits in the URL.
MCPHub: check the version on the hub's dashboard or with npm ls @samanhappy/mcphub, or the image tag you deployed. Then list the hub's users: any non-admin account on 0.12.14 or earlier could have run commands. Look through your server list for entries you did not create, especially any whose command is a shell (/bin/sh, bash) or that call curl or wget, and look in your web logs for POST or PUT requests to /api/servers from non-admin users.
ISPConfig: check the version your panel reports. Anything below 3.2.2 that has not had the 3115_sqlinjection or 321_sqlinjection patch applied is affected. Check that the panel port is not open to the whole internet.
Indicators
None published in the sources we read. The GitHub advisories, NVD, ISPConfig's notice and VulnCheck's catalog entries list no attacker IP addresses, domains or hashes, so we added nothing to the feed. Our "Am I affected?" site checker does not apply: these are server-side flaws, not scripts loaded on a homepage.
Sources
GitHub advisory GHSA-jpw9-pfvf-9f58 (MCP Python SDK): https://github.com/modelcontextprotocol/python-sdk/security/advisories/GHSA-jpw9-pfvf-9f58
MCP Python SDK release 1.27.2: https://github.com/modelcontextprotocol/python-sdk/releases/tag/v1.27.2
NVD, CVE-2026-52869: https://nvd.nist.gov/vuln/detail/CVE-2026-52869
GitHub advisory GHSA-mx89-jjx9-gjr8 (MCPHub): https://github.com/samanhappy/mcphub/security/advisories/GHSA-mx89-jjx9-gjr8
MCPHub release 0.12.15: https://github.com/samanhappy/mcphub/releases/tag/v0.12.15
NVD, CVE-2026-79748: https://nvd.nist.gov/vuln/detail/CVE-2026-79748
ISPConfig 3.2.2 release notice (January 4, 2021): https://www.ispconfig.org/blog/ispconfig-3-2-2-released-important-security-update/
NVD, CVE-2021-3021: https://nvd.nist.gov/vuln/detail/CVE-2021-3021
VulnCheck KEV entries for all three (date_added November 20, 2026, canary-flagged), read through VulnCheck's API: https://vulncheck.com/kev
CISA KEV JSON (checked November 20, 2026: none of the three listed)
Get exploited-bug coverage and a blocklist you can pull for free: register for the DugganUSA feed (docs at docs.dugganusa.com).
Every indicator in this post is in the feed. Free.
1.58M+ IOCs, STIX 2.1 / TAXII, 88% novel vs ThreatFox, exploited-CVE leads ahead of CISA. No credit card — a free API key in 30 seconds, and you can audit every claim above against the live endpoints.



Comments