CVE-2026-114713: What the UnrealIRCd Bug Means for You
CISA put CVE-2026-114713 on the exploited list. UnrealIRCd before 6.2.7.1 lets an ordinary user on a linked network run server commands. Upgrade or patch.
By Muhammad Saif and Aashan Khan · Techtribe co-founders
CVE-2026-114713 is a command-injection bug in UnrealIRCd, and CISA has already put it on the list of vulnerabilities known to be exploited. The National Vulnerability Database describes it this way: 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. On a network with two or more linked servers, a user with no special privileges can inject server-to-server commands that run with server authority. MITRE scored it 8.6, high, with the vector it published: network, low complexity, no privileges, no user interaction, changed scope, high integrity impact. NVD had not yet published its own score when this was written. The CISA date is the one to put on a calendar. The agency added the CVE on November 13, 2026, with a remediation due date of November 16, 2026, for federal agencies under the current BOD guidance. Everyone else does not get a free week because they are not federal. Known exploited means someone is already using it.
The vendor's account is blunter than the CVE text. In the UnrealIRCd 6.2.7.1 announcement, the project says an internal audit, with the help of AI, found several security issues, and that the critical one lets a normal user run commands with server or IRC-operator authority. The examples it gives are operator actions such as disconnecting users or banning them network-wide. It calls this the most critical issue in UnrealIRCd in about fifteen years, and it says it was already seeing exploitation on the internet. That sentence is why this is not a "read the advisory someday" bug. The fix is 6.2.7.1, or the hot-patch bundle the project shipped for the 6.0.0 through 6.2.7 line. This article does not repeat exploit strings, and it does not copy the temporary filter from the forum into a recipe. The forum post is the procedure. Use it on the server you operate.
If the last time you touched an IRC daemon was a side project, a game community, or a bot an agent scaffolded because a tutorial said so, you are the reader. The same habit that misses a PostgreSQL minor will miss this version string. The November PostgreSQL release is the parallel case in what that security release required: a version you can print, a client tool that matches, and a restore you have actually done. UnrealIRCd is the same discipline on a service people forget they still run.
What the bug actually crosses
IRC networks are rarely one process. Operators link servers so a channel on one machine is a channel on the others. The messages between those servers are trusted in a way a client's messages are not. A person in a channel should not be able to speak as the server. Labeled response is an IRCv3 feature that lets a client attach a label to a command so the matching reply can be identified. The flaw is that the label was echoed back into server traffic without being escaped. A label is data. Once it is echoed into a channel that servers parse as commands, data becomes a command, and the command runs as the server.
That is a familiar class, CWE-140, improper neutralization of delimiters, wearing an IRC costume. The condition the CVE states is the one to test in your head before you decide you are safe: two or more linked servers. A single UnrealIRCd process with no link is not the configuration the CVE names. The moment you add a second server, a test box, or a services host that speaks the server protocol, you are in the configuration that is. People add that second server for redundancy and then forget the link is the trust boundary. The bug lives on the boundary.
Impact is integrity of the network, not a silent database dump. With server authority, an ordinary account can do what an operator can do: remove users, ban them, and otherwise abuse the commands the vendor grouped under "many more bad things." For a community, that is a takeover. For a company that parked a support channel or an incident bridge on IRC, it is a way to kick the people who are trying to respond and to impersonate the server while they do it. Confidentiality of a channel log is a separate question the CVE score does not center. Do not argue yourself into delay because the CVSS line says integrity rather than confidentiality. CISA's catalog is not a score debate. It is evidence of exploitation.
The same release fixes more than the critical issue. The vendor lists nine, including a way for a normal user to reveal another user's IP address, a second IP leak, hangs, crashes, a counting error that can make the server tell every new connection it is full until restart, a cap on bans set by services, a TLS renegotiation flood, and a crash in whowasdb. The hot-patch and 6.2.7.1 are how you get all nine. Stopping at "we handled the critical CVE" and leaving the crashes for later is how a half-patched daemon stays on the internet. One maintenance window, the whole bundle.
UnrealIRCd 5 is called out separately. The vendor says it is similarly affected by the critical issue, and that version 5 has been end of life since July 1, 2023. There is no honorable reason to still be on it. The workaround the project mentions for version 5 is to stop loading the labeled-response module and restart, and it says that is not a substitute for leaving the branch. Versions older than that were not fully checked. If a container image says 4.x or 3.2.x, the honest fix is to retire the container, not to hunt a hot-patch that was never promised. End of life means the people who would have written the patch have stopped.
What CISA's due date is for
CISA's November 13 alert added two vulnerabilities. CVE-2026-114713 is the UnrealIRCd delimiter bug. The other is CVE-2026-123039, a Check Point permission-assignment issue. They share a date, not a product. Do not let a patch meeting about firewalls swallow the IRC daemon, or the reverse.
The catalog entry's required action is to apply the vendor's mitigations, follow the current BOD guidance on prioritizing updates and on forensics triage, and, if you cannot mitigate, stop using the product. Federal agencies had a due date of November 16, 2026. Three days is the tell. CISA does not set a three-day clock on a theoretical bug. The vendor had already said exploitation was underway. Your internal SLA of thirty days for "high" was written for bugs nobody was using yet. This one is already in the catalog. If you have a policy that says KEV entries jump the queue, this is the test of whether anyone reads the policy.
Forensics triage, in the BOD sense, means you do not only patch. You ask whether the server was used before the patch landed. On IRC that question is concrete. Were unknown operator actions taken? Did users get removed or banned without an operator doing it? Did the log show commands that arrived as if the server had issued them? A patched daemon with a week of unexplained kicks is a patched daemon you still need to understand. Patching stops the next use. It does not explain the last one.
GitHub's advisory GHSA-237f-p9h6-wq4r restates the NVD description. Use it as a second URL when a scanner wants an advisory ID. Use the vendor forum when you want the fix. Scanners that flag "UnrealIRCd" without a version will nag forever. Scanners that never look at the host because it is not in the application repo will never nag at all. The second failure is the common one for this CVE.
How to prove you are fixed
Prove it the way you prove a database minor: print the version, from the running process, not from a Dockerfile you hope is deployed.
On a server you administer, the vendor's upgrade path for Unix-like systems is the project's own unrealircd upgrade command, followed by a restart if you are not hot-patching. Windows builds, and any hot-patch that fails, take the 6.2.7.1 tarball from the project site. The hot-patch the forum documents is a bundle name you run on each server in the 6.0.0 to 6.2.7 range. After it runs, you still have to check that it took. The project tells operators to inspect the labeled-response module, locally and on remote servers in the network, and to upgrade outright if the check fails. A network is not patched when the server you were logged into is patched. Every linked server has to match, because the bug is in the conversation between them. One old leaf is the bug.
Write down, for each server: hostname, version before, version after, whether you hot-patched or upgraded, and the time. If you applied only a temporary network filter while you scheduled the upgrade, write that too, and write the deadline by which the filter is no longer the fix. A filter that is not on every server at or above the version the vendor requires is not the filter they described. I am not pasting it here. If you need it, you are already on the forum post, and you still owe the upgrade.
Containers and images are where this goes to die. An agent that added a community chat, a status bot, or a "realtime" feature from a tutorial may have committed a compose file with an UnrealIRCd tag that has not moved in two years. Search the repo for the image name and for ircd in config. Search the cloud account for the port IRC actually uses, not only port 443. A daemon that is "internal" and reachable from the office VPN is still a daemon on a linked network if you linked it. Version strings in Slack are not version strings. SELECT version() was the Postgres habit. The IRC habit is to ask the running server and keep the reply next to the ticket.
If you cannot find a version because a hoster hides it, open a ticket and keep the reply, the same way the Postgres note says to do when a dashboard hides the minor. A logo that says IRC is not a patch level. If the hoster cannot tell you whether you are on 6.2.7.1, you do not know if you are exposed, and the KEV clock does not pause for their support queue.
EOL versions get a different sentence in the ticket. "We are on UnrealIRCd 5" is not a status. It is a retirement. Plan the move, take the config you need, and turn the old process off. Do not let an agent "fix" an EOL daemon by regenerating its config in place. That is how you get a new file and the same binary. The vibe-coding security problem is an agent that copies a service from a post and never returns to its advisories. This CVE is one advisory. The next one will not announce itself inside the chat that created the container.
What to look at if you might already have been hit
The vendor said exploitation was happening when it posted. CISA agreed by listing the CVE. You will not get a personalized indicator in this blog, and you should not want one. Look at your own logs for operator-level actions that no operator took, for mass disconnects, for bans that do not match your policy, and for links to servers you did not add. If the server is a community you do not actually need, take it down until it is on 6.2.7.1. Availability of a hobby network is not worth a weekend of arguing with an attacker who already has server authority.
If the IRC server sits next to something that matters, a support workflow, a deploy bot, an alert bridge, assume the channel was not a trustworthy control plane while the daemon was unpatched. Move those notifications to a system you patched on purpose. Do not keep the on-call bridge on a daemon whose version you learned about from a news post.
Rotate operator credentials after the upgrade if you cannot account for who used them. Server authority in the bug means an attacker did not need to steal the operator password to act as the server. They may still have taken credentials if they operated the network for a while. New operator passwords, a review of who is an operator, and a look at services passwords are cheap compared with discovering a second operator you did not appoint.
Back up the configuration before you upgrade, and know how you would restore it. A hot-patch that avoids a restart is attractive because the network stays up. It is not a reason to skip a copy of the config. If the upgrade path restarts and the config does not load, you want the previous file. Test that you can start the new version with that file on a spare host if you have one. A production network with a single server you have never rebuilt is one bad restart away from a longer outage than the patch. The Postgres advice applies in miniature: two copies, one of them proven.
What to do this week
- Find every UnrealIRCd process, image, and host, including side projects and old containers. If the search is empty and you have checked the cloud account as well as the repo, write "none" in the ticket and stop.
- For each process, record the version from the running server. You want 6.2.7.1 or newer, or a hot-patch the vendor's check confirms. "6" is not a match. "Latest" is not a match.
- Hot-patch or upgrade every linked server, not only the one you logged into. Confirm labeled-response on the remote servers the way the project describes. If the hot-patch fails, upgrade and restart.
- If any server is version 5 or older, schedule removal. Do not spend the week tuning an end-of-life branch.
- Read logs for operator actions you did not take, before you call the ticket done. Patching is step one. The due date CISA set assumed you would also ask whether you were already late.
- Move any on-call or deploy signal off an unpatched IRC network today, even if the upgrade is tonight.
- Put the version check next to the other daemons an agent installed and nobody owns. The next KEV entry will not be IRC. It will be whatever still has no owner.
A version string you can paste into the ticket is the whole job. The announcement is not the install. CISA's catalog is not the install either. The install is the binary that is running after you log in.
If this is your app, send the repo.
A 30-minute scope call. We read what is live, name what is dangerous, and quote a fixed scope.
How Techtribe checks a daemon a generator left running
We search the repo and the hosts for the service the advisory names, record the version the process is actually on, and separate a hot-patch from an upgrade that needs a restart. You get that note in writing, plus the end-of-life boxes that should be turned off instead of patched. We do not paste exploit strings into the ticket, and we do not let an agent rewrite the config in the same window and call that the fix.
If the service is one forgotten container, the pass is small. If the product is a pile of generated services and none of them have a version owner, that scope is a Production MVP. The scope call is 30 minutes with a founder, and it is where we say which you are in.
Bring the host list and any IRC, bot, or "realtime" container you remember. If you have never printed the version, say that. It changes the order of the call.
The fixed scope for taking a generated app to production is a page per tool: Lovable, Bolt, v0, and Replit.
Questions
What is CVE-2026-114713?
It is a flaw in the labeled-response module of UnrealIRCd 6.0.0 before 6.2.7.1. The IRCv3 label tag is not escaped when the server echoes it, so a user on a network with two or more linked servers can inject server-to-server commands that run with server authority. CISA added it to the Known Exploited Vulnerabilities catalog on November 13, 2026.
Which versions are affected?
NVD lists UnrealIRCd before 6.2.7.1, and the vendor describes the critical issue in the 6.0.0 line up to but not including 6.2.7.1. The vendor also says UnrealIRCd 5 is affected and has been end of life since July 1, 2023. Older end-of-life branches were not fully checked. If you are not on 6.2.7.1 or the vendor hot-patch, treat the server as exposed.
Is there a fix that avoids a restart?
The UnrealIRCd project says most version 6 networks can hot-patch without a restart, using the bundle it published, and that you should confirm the labeled-response module afterward. If the hot-patch fails, or you are on Windows, upgrade to 6.2.7.1 and restart. End-of-life version 5 does not get that hot-patch. Move off it.
Why does this matter if we do not run IRC?
It matters if any side project, community server, support bot, or old container still does. It also matters as a pattern: a parser bug on a linked service becomes administrative authority, CISA sets a short deadline once exploitation is known, and 'we will get to it' is not a version check.