N034: Exploit code is out for AI-found NGINX flaws on 9 million exposed services

Exploit code is public for AI-found NGINX HTTP/2 flaws, Canonical ships AI-written Ubuntu patches before testing, and a lapsed domain gave researchers a top Hugging Face model's account.

Share

N034 · Thursday 19 November 2026 · 5-min read · Coverage: 18 Nov 4 AM PT to 19 Nov 4 AM PT

Jump to: Outside agents today · Incidents · Agent threats & vulnerabilities · Policy, protocols & compliance · Architecture & harnesses · Models & infrastructure · Defenses, tools & GitHub · Pro deep dives

Today in 30 seconds

  • Rapid7 and Censys published exploit code within a day of F5 fixing two NGINX HTTP/2 bugs found by Wiz's AI scanner; Censys counts about 9 million exposed services.
  • Canonical now lets Ubuntu Pro customers install AI-agent-written security patches before its engineers finish testing them.
  • Researchers took over the Hugging Face account behind a model with 13 million monthly downloads by registering its lapsed domain.

Outside agents today

  • Apache Roller: Apache says attackers are exploiting CVE-2026-129448 (CVSS 9.8), an unauthenticated code-execution flaw in versions 4.0.0 to 6.1.4 (ASF advisory); upgrade to 6.1.6 or remove the XML-RPC servlet mapping (actively exploited).
  • F5 BIG-IP: Six CVSS 9.8 overflows let unauthenticated traffic crash TMM when certain iRules or APM policies are used (F5 K000164058); only engineering hotfixes exist.

Top story

AI-found NGINX HTTP/2 flaws now have public exploit code

18–19 Nov 2026 · F5 K000164058 · Wiz · Vulnerability · CVE-2026-123633, CVE-2026-123007 · Relevance 9 · Editor pick

Widely used web server software has two bugs that let anyone on the internet run code on it when HTTP/2 is on, and an AI scanner found them. Wiz Atlas, Wiz's AI code-scanning harness, found a use-after-free in HTTP/2 connection handling (CVE-2026-123633, mainly proxies) and a send-window underflow (CVE-2026-123007, also static-file servers), both CVSS 4.0 9.2. F5 fixed them 18 Nov; Rapid7 and Censys then published proof-of-concept exploits (SecurityWeek). No exploitation is reported.

For leaders: Your patch window for internet-facing servers is now days; confirm who owns emergency NGINX upgrades, including copies inside appliances and ingress controllers.

For builders: Upgrade to NGINX 1.30.7 or 1.31.9 (Plus 37.0.8.1 or 37.3.0.1); where you can't today, set http2 off;, remove http2 from listen lines, reload and confirm old workers exited.

Pro deep dive ↓

Incidents

Update: Uber says its AI vendor's breach exposed 149,346 drivers' license records

Filed 10 Nov, reported 17 Nov 2026 · Maine AG filing · Incident · no CVE · Relevance 8 · US

Hackers who broke into a company running AI services for Uber took about 149,000 drivers' names, license numbers, birth dates and addresses. TechCrunch names the vendor as Fireworks AI, whose breach also touched Notion and Cursor data (N031).

Do this: Ask each AI inference vendor what data it keeps beyond the request, where, and for how long.

Pro deep dive ↓

Agent threats & vulnerabilities

17 Nov 2026 · BleepingComputer (secondary) · Vulnerability · no CVE · Relevance 7 · Global

Researchers bought an expired domain listed in a popular AI voice model's files and used it to take over the publisher's account, which could have pushed altered models to millions of users. Alon Hertz's team reset Kokoro's Hugging Face password through the domain and tested poisoned weights only locally; Hugging Face had not responded by 17 Nov.

Do this: Load models by commit hash, not repo name, and alert on publisher changes.

Pro deep dive ↓

Policy, protocols & compliance

UK MPs want a regulator that can halt AI training and suspend deployed services

18 Nov 2026 · The Register on the Commons committee report "Guardrails for Growth: Making AI Safe" · Proposed · UK · Relevance 7

British lawmakers want a regulator that can stop risky AI models mid-development and switch off services already in use. The committee's proposed AI Safety and Security Bill adds powers to block deployments and mandatory incident reporting, including near misses; the government hasn't committed.

What it means for US companies: US AI providers selling in the UK would face the same suspension and reporting powers.

Do this: Start logging near-miss agent incidents now.

Also:

Architecture & harnesses

Canonical ships AI-agent-written Ubuntu security patches before human testing ends

12 Nov 2026 (reported 18 Nov) · Canonical · Release · no CVE · Relevance 7 · Capability shift

Ubuntu's maker now offers security fixes that AI agents wrote and other AI models approved, installable before engineers finish testing. In Canonical's Surge pipeline, several models write competing backports and judge models must agree; unclear cases go to a human, and fixes ship only via the opt-in Ubuntu Pro Edge archive.

New risk for leaders: The review gate moves from the vendor to your change process; trusting the distro no longer covers these packages.

Do this: Keep Edge off production; test only on canary hosts.

Pro deep dive ↓

Models & infrastructure

Also:

Defenses, tools & GitHub

Also:

One concept

Expired-domain account takeover

When the domain behind an account's email lapses, whoever re-registers it can receive password resets and take over the account. For a model publisher, that means control of what every downstream user downloads next.

Basic mitigation: Pin models and packages by hash.

Full glossary


AgentSec Brief is published by AM Technology Solutions. Researched and drafted with AI agents, verified against primary sources, and reviewed before publication. Informational only; not legal, compliance or security advice. Verify before acting.

Pro members get the full deep dives with attack diagrams, detection queries and the fix in order, plus the vulnerability tracker, compliance calendar, open-source picks and watch list. Upgrade at https://www.agentsecbrief.com/#/portal/signup