UnrealIRCd, the most widely deployed open-source IRC server, has a command-injection flaw that is being exploited. CVE-2026-114713 (CVSS 3.1: 8.6) lets an ordinary connected user, with no oper privileges, smuggle protocol lines into the server’s own replies. Those lines then run with server or IRCop authority. The project shipped UnrealIRCd 6.2.7.1 and a hot-patch on November 7, 2026. Exploitation attempts started about a day after disclosure. The maintainer calls it the worst UnrealIRCd bug in roughly 15 years. CISA added it to the Known Exploited Vulnerabilities catalog on November 13 and set a federal remediation deadline of November 16.

IRC can look like a legacy concern. It isn’t one for infrastructure teams. UnrealIRCd still sits behind open-source project chat, gaming and hobby networks, internal ops channels, and a large share of botnet command-and-control. A network-wide GLINE or mass KILL from a fake “server” is a real outage. Linked networks spread the damage to every server in the mesh.

Technical Details

The bug is in the labeled-response module, which implements the IRCv3 labeled-response capability. A client that negotiates this capability can attach a label message tag to any command, for example @label=abc123 PRIVMSG #chan :hi. The server echoes the label back on its reply so the client can match responses to requests.

In versions 6.0.0 through 6.2.7.0, the server copied the client’s label value into the reply without properly escaping it. IRC is a line-oriented protocol, and message-tag values have their own escaping rules (\s, \:, \r, \n, \\). A label containing raw or improperly decoded CR/LF breaks the reply into several protocol lines. The attacker controls everything after the break. Because the injected lines go out in the server’s own context, including over server-to-server links, they are processed with server privileges.

According to the maintainer’s oss-security post, a normal user can use this to run commands normally limited to IRC operators or servers, including:

  • KILL: disconnect arbitrary users, including opers
  • GLINE / ZLINE: place network-wide bans
  • other server-originated state changes that linked servers accept as authoritative

There is no authentication barrier beyond being able to connect and negotiate the capability, and most public networks allow both by default. A public proof-of-concept is circulating, so the barrier to exploitation is effectively zero.

Affected and Fixed Versions

BranchStatus
UnrealIRCd 6.0.0 – 6.2.7.0Vulnerable
UnrealIRCd 6.2.7.1Fixed. Release also addresses eight other issues
UnrealIRCd 5.x and earlierEnd of life. Vulnerable where labeled-response is loaded, and no fix is coming

Distro packages lag upstream. Red Hat’s tracker notes that EPEL 9 still ships unrealircd-6.1.10-1.el9, which is affected, and an update to 6.2.7.1 is pending. Check Gentoo, Debian/Ubuntu PPAs, container images, and any vendored builds separately. Don’t assume your package manager has caught up.

Impact

  • Network disruption: attackers can mass-kill users and opers, G-line entire IP ranges, and lock legitimate staff out.
  • Trust abuse across links: in a multi-server network, one unpatched leaf is enough. Injected server-originated lines are trusted by peers.
  • Follow-on compromise: this is not code execution on the host. But attackers who can impersonate server authority can take over channels, harvest traffic, and target ops staff for social engineering. That matters most where IRC is used for internal operations or incident coordination.

Mitigation

  1. Upgrade to UnrealIRCd 6.2.7.1 or later now. The project also published a hot-patch for 6.x if you can’t do a full rebuild right away. Apply it on every linked server, not just hubs.
  2. If you can’t patch immediately, disable the vulnerable module. Add blacklist-module "labeled-response"; to your config and restart (a rehash does not unload it). This breaks IRCv3 clients that rely on labeled replies. It does not affect core chat.
  3. UnrealIRCd 5 operators: blacklist the module as above, then plan the migration to 6.x. There will be no fix for 5.
  4. Hunt for abuse. Review logs from November 7 onward for unexpected KILL, GLINE, ZLINE, or SHUN events, unexplained oper or mode changes, and client lines whose label tags carry escaped or raw CR/LF sequences. Remove any bans you didn’t set.
  5. Federal agencies under BOD 22-01 had until November 16 to remediate. CISA’s entry also asks for forensic triage of exposed instances.

References