Two stories landed in the same fortnight that this site would normally run as two posts, and they belong together: in one, an automated agent earned a line in a vendor's acknowledgements for finding a critical web-server bug; in the other, a hosting company is publishing its own incident page explaining that an attacker took over the MCP server that the company had connected to its hosting control plane. Agents on both ends — the finder and the door. The homework from both is unusually short and unusually overdue: two commands on every box that runs Apache, and one inventory question about every MCP endpoint anyone in your organization runs or exposes.
The Apache bug: a race in mod_http2, reported by a machine¶
Upstream numbered it CVE-2026-119618, ranked it critical on Apache's own impact scale, and described it precisely: a concurrency bug — "Concurrent Execution using Shared Resource with Improper Synchronization" — in assign_headers() in mod_http2, the module that speaks HTTP/2, with heap corruption that can lead to remote code execution. It affects Apache HTTP Server 2.4.55 through 2.4.70 — which is every currently packaged 2.4.x — and, separately, the standalone mod_h2 project on GitHub in versions 2.0.8 through 2.0.42. The fix is commit 824beab, and the release landed November 11, 2026.
The acknowledgements section is the part this site keeps a ledger on. The finders: "Anthropic's OSS Scanner, as vulnerability ANT-2026-MAPKT6MZ" — the vendor is crediting a machine with a machine-issued tracker ID — plus Anthony Weems and Simon Scannell of Google Cloud Vulnerability Research. Anthropic's report reached the Apache security team on October 26; Google's on November 3; fix and release came November 11. Sixteen days from report to patched tarball for a critical in the most-deployed web server on the internet is a good number. This ledger now reads: the IPv4 handshake race found by Anthropic's research model, fixed upstream October 15; the musl/OpenSSH password bypass, found by a human with a 12-build matrix; and now a second shared-state race in mod_http2 this quarter — CVE-2026-57941 (a use-after-free via shared session->bbtmp re-entrancy, fixed October 1) is followed by this one. Two race-class bugs in the HTTP/2 module in two consecutive releases is a pattern, and mod_http2 has now produced criticals in May, October, and November by heise's count.
2.4.71 carries two companions. CVE-2026-106023 (important): a NULL pointer dereference in mod_md during request processing — child process crash, denial of service, affects 2.4.30 through 2.4.70. And CVE-2026-109163 (low): a one-byte stack buffer overwrite in mod_vhost_alias, listed by Apache as "after CVE-2026-63292" — meaning the October 1 fix for vhost_alias did not hold: 2.4.69 and 2.4.70, the very releases that contained that fix, are affected again. The same "fixes that did not hold" family this site documented in Keycloak 26.8.2: a fixed-in-N advisory is a claim about today, not a property of the release number.
What you should actually do about it¶
Check first — mod_http2 is not enabled by default on Debian-family packages, so a large share of self-hosted boxes can close this entirely with a look:
apache2ctl -M | grep http2— is the module even loaded? If not, you are done.grep -rE '^\s*Protocols' /etc/apache2/— what do your vhosts negotiate? Anything still listingh2orh2con an internet-facing vhost is the exposed surface.- Version:
apache2ctl -v. Anything 2.4.55–2.4.70 with the module loaded is in range;mod_mdusers (2.4.30+) sit in range for the crash bug regardless ofhttp2.
If you can upgrade, that is the fix: upstream 2.4.71, or your distro's build of it. Ubuntu shipped USN-9090-1 on November 12 — one day after the upstream release. Debian's security tracker carries the CVE with the upstream fix commit recorded; run the tracker check before assuming a point release exists, and pin the release date of whatever you install, because the 2.4.69 lesson applies in both directions.
If you cannot upgrade today, the fallback Apache names is one line, and heise relays it as the honest interim: Protocols http/1.1 in the httpd config. That drops h2 and h2c negotiation entirely — HTTP/1.1 only, which is fine for a community site, a fire hall's server, a small business vhost, and materially slower only for very chatty browser traffic. If your concern is the mod_md crash instead, that is load-bearing for ACME management; upgrade for that one, there is no protocol switch that avoids a module you need.
If you run the standalone mod_h2 rather than Apache's bundled copy, the affected range is 2.0.8–2.0.42; the advisory names the range but not a fixed release, so the honest move is to watch the upstream tracker and not to assume the bundled-module fix shipped in your build.
Mittwald: the week's other direction¶
The German hosting provider Mittwald — tens of thousands of small-business and agency sites — published its own account, and it matters more than the headline. From the company's status page and its incident FAQ, quotes translated by this site:
- October 27: unauthorized access to the MCP server began. Per the FAQ: "Der MCP-Server ist eine Schnittstelle, über die Kunden ihr mStudio mit KI-Assistenten verbinden konnten" — the MCP server was an interface through which customers could connect their mStudio hosting console with AI assistants. "Der Server war ab dem 27.10. übernommen. Jede Nutzung danach gilt als potenziell kompromittiert." — the server was taken over from October 27; every use after that counts as potentially compromised.
- November 4: the company's own analysis detected the intrusion. When asked why it took a week, the FAQ's answer: the attacker deliberately camouflaged his preparations "using actions that looked like completely normal user behaviour."
- November 5: lock-down at 11:34 CET — mStudio logins, all SSH, the platform extensions. At 17:49 CET the status update names it: "unbefugten Zugriff auf das mStudio über unseren MCP-Server festgestellt" — unauthorized access to mStudio via the MCP server. Session tokens of users of that interface were read out; the FAQ is explicit that no passwords were taken, and that a token "erlaubte Zugriff auf dein mStudio mit deinen Rechten" — allowed access to your mStudio with your own rights, i.e. every project that user could reach.
- November 7: mStudio and SSH restored, passkeys introduced, and a clarification: the customer center and the websites managed in it "were at no point affected by the security incident." Preemptively locked customer projects — a small number per the company, 56 per one report — stayed locked pending scans.
- November 12, 02:51–06:52 CET: the attackers resumed. Per the status page: "Bislang ist es den Angreifern gelungen, eine Sicherheitslücke auszunutzen und unsere Unternehmenswebseite zu kompromittieren" — a second vulnerability, this one compromising the corporate website, which the company pulled and replaced with its status page; "Zugriff auf unsere internen Systeme konnten wir bislang nicht feststellen" — no access to internal systems detected so far. The characterization: "professionelle Akteure ... die automatisierte KI-Systeme nutzen" — professional actors using automated AI systems. Kundencenter, mStudio and SSH went offline again, preemptively, this time until after individual project scans complete; re-open was targeted for November 13, 11:00 CET. Hosted customer websites stayed up through both waves. heise's update carries a defacement claim from a whistleblower alongside the company's denial — "Die Unternehmenswebsite hat mittwald vorsorglich selbst vom Netz genommen ... Eine Übernahme oder ein Defacement der Website durch Angreifer fand nicht statt" — a two-source dispute this site records rather than resolves.
Two structural points survive every open question. First: the entry point was not a PHP bug, a leaked password, or a zero-day in the hypervisor — it was a new, AI-era, always-on API endpoint that translated natural language into privileged console operations, authenticated by whoever held a session. Second: the blast radius was bounded by rights that were themselves the data — the tokens carried "your own rights," which is exactly why per-project forensics is taking the company this long.
This is the first breach this site has found on the record where the named entry point was the MCP server itself, and it arrives while self-hosting communities are actively wiring the same pattern — an agent gateway in front of real infrastructure — onto smaller consoles. The lesson is not "MCP is insecure." The lesson is that an MCP endpoint is an admin API with a chat interface bolted on, and it inherits admin-API obligations: scope the credentials it holds, know which sessions it can move laterally between, log its tool calls like an audit trail, and keep it off the public internet unless someone can say out loud why it belongs there. The company's own remediation — server offline indefinitely, CLI kept as the automated alternative — is itself the telling sentence: the agent interface was not load-bearing enough to bring back before it could afford not to be.
There is an echo with this site's php2shell coverage: this is the same hosting provider that preemptively suspended customer projects during the October PHP crisis, and three weeks later it is doing the same thing to its own console. Twice in a month, the correct move for a host this size was to turn things off and do forensics in the dark. The php2shell week's lesson was that turning things off is a government-scale precedent; Mittwald is the enterprise-hosting version — availability traded knowingly for containment, customers informed in writing each time.
The inventory question, written down¶
For the Apache side, the two commands above settle it on any box in ten minutes. For the MCP side, the question Mittwald's incident answers is one every operator who has stood up an agent endpoint — or is about to — should be able to answer on paper:
- What exists? List every MCP gateway, agent endpoint, or "connect your AI assistant" feature you run or allow: hosted SaaS consoles, internal tool wrappers, self-hosted servers bound to
localhost:portfor one person's IDE and forgotten. - What does a session leak? For each: what credentials or tokens does the endpoint hold, and what does one stolen session reach — with whose rights? This is the number Mittwald had to discover under fire, and the one whose size made the difference between an incident and a crisis.
- Who can reach it? Public internet, allow-list, VPN-only, or loopback. The company that breached itself into a fortnight of SSH downtime ran it on the public internet for customers by design; that was a product decision, and this incident is its cost line. Your community's agent wrapper probably has no such justification.
- Where is the log? Tool-call audit trail with retention, or console output nobody keeps? If there is no log, an incident is a story: the attacker camouflaged his preparations using actions that looked like completely normal user behaviour is what a week undetected sounds like when the log is quiet.
- What is the off switch? Mittwald is demonstrating the off switch is real: server offline indefinitely,
CLIretained as the automated path. If your agent endpoint cannot be shut without stopping work, it is load-bearing, and load-bearing components deserve the treatment the AI era has been giving everything else: version-pinned, logged, and assumed to break.
The customer homework Mittwald put in writing is worth copying verbatim into your own incident runbook, because it is the generic post-token-theft checklist: change the password, enable two-factor, audit every API key, project token and SSH key including SSH users you did not create, rotate every third-party credential stored inside your projects (databases, mail, external APIs), tell your users. The compromise indicators the company named — unknown SSH users, unrecognized files in the project directory, API keys you never created — are the same three things to grep for on any system where a session token could have moved sideways.
The honest column¶
- No CVSS has been published for CVE-2026-119618 as of this site's check on November 13; "critical" is Apache's own impact rating. The advisory row on the distro trackers is thin by design this early. Treat severity as claimed, not scored.
- No public PoC, and no reported exploitation of the Apache bug as of publication. Kevin Beaumont's assessment relayed by heise — hard to exploit because of ASLR — is a reasoned estimate, not a proof. The race-to-RCE chain is the vendor's characterization of its own bug. None of that is a reason to leave h2 enabled on a box you could patch in five minutes; it is a reason not to panic about boxes you will patch this weekend.
- The mod_h2 fixed version is unpublished in the advisory. Standalone users: verify, do not assume.
- Mittwald's technical specifics are the company's own: the October 27 takeover date, the token mechanism, the "no passwords" statement, the "no internal-system access" claim, and the camouflage explanation all come from its status page and FAQ, which have published no packet-level detail. The 56-projects figure is a single third-party report; the company says "a small part of projects." The second vulnerability's identity is unnamed. The whistleblower's defacement claim and the company's denial are recorded above and left unresolved. The company has said it will publish its learnings "openly and transparently" when its analysis is complete — that promise is on the record now, and this site will grade it against the standard AWS's sovereignty drill already set: claims go in the artifacts.
- The fix's durability is unproven by time: October's 2.4.69 vhost_alias fix did not hold for the two releases that carried it. "Fixed in 2.4.71" is the current state of a moving target, and the same week Keycloak shipped 26.8.2 with three fixes that did not hold.
- What survives all caveats: two commands, one line, and one inventory — the patch,
Protocols http/1.1if you cannot, and a written answer for every agent endpoint about what a stolen session reaches. Agents found the bug this week and were the door this week. Neither fact changes what the boring middle has always been.
Citations verified live on 2026-11-13: upstream advisory, USN-9090-1, Debian tracker, heise (both stories), Mittwald status page and FAQ — 1 2 3 4 5 6 7 8 9 10 11. German quotations translated by this site. Content WTFPL, code AGPLv3.