Two entries, Friday's alert, one November 16 deadline for federal agencies: CVE-2026-114713 (UnrealIRCd) and CVE-2026-123039 (Check Point multiple products). This site pulled the catalog JSON and counted the week: four entries November 9, six on November 10, two on November 13, nothing in between — so the November 9 batch post and Friday's two entries close the ledger for November as of this writing.

UnrealIRCd: the label tag, the linked network, and any user

The NVD description is precise and short enough to quote whole:

The labeled-response module in UnrealIRCd 6.0.0 before 6.2.7.1 does not escape the IRCv3 "label" message tag when echoing it into replies, which allows any user on a network with two or more linked servers to inject server-to-server commands that are executed with server authority.

IRC is a text protocol; message tags are newline-tolerant fields in the spec, so if the code echoes an unescaped tag back into server-to-server traffic, the injected newline becomes a new line of protocol — a complete command, authored by any connected user — that the receiving server executes with server authority. The maintainer's oss-security post puts the practical result flatly: any ordinary user can KILL and GLINE "and do many more bad things." He calls it the most critical issue found in UnrealIRCd in the past 15 years, and nothing in the CVSS 8.6 argument changes the description. NVD's score (AV:N/AC:L/PR:N/UI:N/S:C, integrity high, CWE-140) even carries a confidentiality rating of none — a scoring of a control-plane takeover that rates the takeover as no impact on confidentiality. The score travels even when it disagrees with the reality it is supposed to model.

Three scoping facts matter for who is actually exposed:

  • Two or more linked servers. The CVE's own text scopes the injection to networks with server-to-server traffic. A single-box IRC network stands outside the CVE's scoping — and its operators should apply everything below anyway, because the same bundle fixes eight other issues including two that leak a cloaked user's IP address.
  • Versions 6.0.0 through 6.2.7 carry the bug. The fix is 6.2.7.1, released Saturday November 7.
  • End-of-life lines are not spared. UnrealIRCd 5 — EOL since July 2023 — is similar, and the maintainer says flatly that nobody should be running it. Older 3.2.x and 4.x were not checked; his words are "you are not safe at all."

The nine issues fixed in the bundle, from the advisory: the critical label-escaping command injection (the only one with a CVE number); two decloak/IP-disclosure leaks; a hang in blacklist rechecks; a services-triggerable crash via CHGIDENT that a normal user can induce; a counting error that marks the server "full" to every new connect until restart; a cap on services' ban count; a TLS-renegotiation flood; and a whowasdb crash.

The fix shipped Saturday. Scans began Sunday. KEV landed Friday.

The timeline is the part worth copying into your own incident playbook:

  • Wednesday, November 4 — pre-announcement on the mailing list, forum, main site, Mastodon and X: a critical issue exists, get ready.
  • Saturday, November 7 — UnrealIRCd 6.2.7.1 and the hot-patch bundle ship, and the maintainer sends "almost 1000" automated emails to the admin contacts declared in each network's MOTD and ADMIN lines. A second wave goes out two hours later.
  • Sunday, November 8 — the CVE record publishes, the oss-security advisory goes out — and the maintainer reports the abuse attempts starting, then spends the day connecting to every vulnerable network with more than 100 users to hand-deliver the link in person.
  • Friday, November 13 — CISA certifies exploitation and adds the entry to the catalog.

That is a six-day fix-to-KEV gap — against the 144 days (GPUStack) and 69 days (FreeSWITCH) this site measured the same week. Read the comparison correctly: the six days are not a win, they are a different shape of failure. Exploitation did not wait for the gap to close — it started the day after disclosure, because a working public description of a root-the-network bug is a working exploit for everyone who was waiting on one. The head start went to roughly a thousand emailed admins; every operator who found out through the catalog on Friday learned about their already-patched bug one day later, and about their unpatched window exactly when the scans were already underway.

The reason the fix could ship with zero downtime is hot-patching, a mechanism UnrealIRCd has run for twenty years — the first one shipped January 2005, and by the vendor's tally more than 80% of its vulnerabilities have been hot-patchable since. ./unrealircd hot-patch bundle-20261107 downloads a GPG-verified patch, dry-runs it, applies, recompiles, and rehashes; the details below.

The ledger: the maintainer and the AI

This site keeps a running ledger of AI-era discovery — the Anthropic scanner that found the Apache mod_http2 race, the solo researcher and his 12-build matrix behind the musl star bug, Binarly against Zyxel's firmware that did not work. Every previous entry had a third party digging into someone else's code. This one is the maintainer chapter, and it reads the old way with new tools: Bram Matthys has audited his own C codebase since 2001 — static analysis era, fuzzing era — and since February spent about a hundred hours of that with AI added to the loop. In his words, "especially the last month things have been wild," and this batch — nine issues in one internal audit — is what came out of it.

His conclusion is the quiet-daemon line, and it deserves to be quoted intact because it is the sharpest warning any of this month's disclosures has produced:

I genuinely think that, if you are running any network daemon that has not seen any major security fixes in the past couple of months, then nobody has been poking at that daemon with AI. Or some bad guy did but did not report back.

Every community-network story this site has written this quarter is that sentence wearing different clothes: 996 Zyxel switches emptied across 48 countries, FreeSWITCH PBXs reachable by any registered account, routers taken without passwords because a community network is the target that does not update. The Zyxel fleet's problem was not that someone searched it — it was that nobody had searched it until it was worth money to someone. The inventory question now has a clean yes-or-no form your logs can answer: when did this daemon last ship a security fix? A daemon with no fix in months has not survived scrutiny. It has avoided it.

Check Point: a root daemon on a world-writable socket, briefly

CVE-2026-123039 sits in the batch for the pattern rather than the product. On Check Point Gaia OS, the DynamicCLI daemon runs as root and listens on the local socket /tmp/dynamic_socket without authenticating connecting processes; per the NVD row (published November 12, still Undergoing Analysis; CVSS 7.8 local — vendor score pending NVD) an authenticated local user sends a command to the socket and it executes as root, across firewall, security gateway, and management products. Check Point's advisory is sk1000362. Community networks do not run Check Point; most will never read the SK. The class, though, showed up twice in one catalog week: a privileged listener that never authenticates its peers erases the local trust boundary, exactly like GPUStack's world-readable /var/lib/gpustack credentials let any local user forge admin tokens. Whether you run the appliance or the inference cluster, the check is the same one the BOD implies: list the local-privileged listeners on your box and ask which ones would execute commands from any process that can connect.

The honest column

  • Exploitation evidence is the vendor's own sighting. The catalog's notes link straight to the UnrealIRCd forum thread; there is no public PoC, no GreyNoise-style telemetry page for IRC yet, and no attack-volume numbers. "Bad guys are exploiting it right now" is a maintainer watching his networks get hit, which is real evidence — but it is a census of one.
  • Only one of the nine fixed issues carries a CVE identifier. The other eight are findable only by reading the advisory. If the forum thread ever moves or goes down, the catalog's reference rots with it — the documentation trail for a critical entry is one phpBB page.
  • The scoring quirk cuts both ways and both directions get named: CVSS 8.6 with C:N/A:N understates a bug whose practical result is "any user rewrites your network's control plane"; NVD's Check Point row is one day old and still Undergoing Analysis, so its 7.8 is effectively a vendor score wearing an NVD URL.
  • The spamfilter workaround protects all linked servers — but only if every server on the network is already 6.0.4 or newer. A version dependency buried in step one of an emergency procedure is exactly the kind of thing a mixed-version network misses on a Sunday.
  • The hot-patch requires a working build environment; the FAQ exists because that is not automatic — including the distro-upgrade case that can break the build. The verification step is not optional hygiene; it is the only way to know the patch actually landed.
  • The AI-audit chapter is a self-report. No methodology was published, and none needs to be — the nine fixes are in the source for anyone to verify. But "AI found it" should be read as "a maintainer with twenty years of context drove the audit," not as a tool you point at code to get the same output.

For your deployment

  • IRC with more than one server, on any UnrealIRCd 6.0.0–6.2.7: apply the workaround first, as an IRCOp, in one command — /SPAMFILTER add -regex T block - Invalid_label ^label=(?![^\r\n;\x20]*\z) — then on every server run ./unrealircd hot-patch bundle-20261107 (or upgrade to 6.2.7.1 and restart). Verify each box individually: ./unrealircd module-status | grep labeled-response should print 5.0-hot-patch-bundle-20261107 — a bare 5.0 means the hot-patch is not applied — and from IRC, MODULE -all remote.server.name reaches the servers you cannot shell into.
  • UnrealIRCd 5: same critical issue, no hot-patch path. Add blacklist-module "labeled-response"; to the config of every server and restart — the vendor underlines that a rehash is not the same operation — then plan the upgrade; the line has been end-of-life since July 2023.
  • Assume-compromised homework, federal deadline or not: logs back through November 7, before the patch date. Spamfilter hits on Invalid_label are the smoking-gun signature that someone tried; audit KILL and GLINE notices, OPER lines, services activity, and anything your network did that you cannot account for. Rotate operator credentials where the logs are ambiguous.
  • Single-server IRC networks: outside the CVE's own ≥2-linked-servers scoping, still patch — the bundle carries the decloak leaks and a user-triggerable crash that do not care how many servers you run.
  • And the standing inventory question: the daemon that answers "when did I last ship a security fix?" with "not in months" is the one the scans are coming for. That applies to the switch in the rack, the PBX in the closet, the chat server nobody remembers installing — and, this week certifies, to the IRCd at the heart of a network's coordination.

Sources: CISA alert, November 13, 2026; CISA KEV catalog JSON (fetched 2026-11-14T02:01-07:00, queried post-add — dateAdded 2026-11-13, dueDate 2026-11-16); UnrealIRCd advisory, forums announcement by Bram Matthys, November 7, updated November 8; oss-security: UnrealIRCd 6.2.7.1 released & hot-patch (CVE-2026-114713), November 8; NVD CVE-2026-114713 (published 2026-11-08, CVSS 3.1 8.6, CVE lastModified 2026-11-14); NVD CVE-2026-123039 (published 2026-11-12, Undergoing Analysis, CVSS 3.1 7.8); Check Point sk1000362; Why hot-patching is awesome, vulnscan.org; this site's November 9 KEV batch post and Zyxel KEV post.