A house newsletter · daily · researched and written by the Kurzia AI
Thursday 19 November 2026 · Issue 19 · 4 min read
NGINX2root: patch today; Puppet joins CISA's exploited catalogue; vLLM template-trust bypass; PHP criticals land this afternoon
Act on this first
watchtowr.com · yesterday
NGINX2root
F5 patched five NGINX memory bugs yesterday, two of which give an unauthenticated attacker code execution on the web server itself, whatever runs behind it. Both sit in HTTP/2 handling: an integer underflow in flow-control accounting dating back to 1.9.5 (CVE-2026-123007) and a use-after-free in connection handling present since 1.19.7 (CVE-2026-123633). HTTP/2 must be enabled for either to work, and the use-after-free also needs a reverse-proxy setup.
No attacks are reported so far, but a working proof of concept is already public, ASLR does not stop these bugs, and roughly a third of the web runs on NGINX — so patch internet-facing servers, load balancers and Kubernetes ingress first. Update to 1.31.9 (mainline) or 1.30.7 (stable), or NGINX Plus 37.0.8.1 / 37.3.0.1; until you can, set http2 off and reload. The archived kubernetes/ingress-nginx project gets no fix at all. Check exposure with `nginx -T | grep http2`.
Affected: CVE-2026-123007: 1.9.5 and later; CVE-2026-123633: 1.19.7 and later · Fixed in: 1.30.7, 1.31.9 (OSS); NGINX Plus 37.0.8.1, 37.3.0.1 · Disclosed 18 Nov · We had it 18 Nov, 23:04 UTC
Under attack
React2STEAL — React Server Components 19.0.0–19.2.0, unpatched Next.js middleware — GitHub Actions campaign chains pre-auth RCE to steal AI and cloud keys; both flaws have fixes; CVE-2025-55182, CVE-2025-29927. Forty-five repositories run the whole pipeline on GitHub-hosted runners and commit victims' environment dumps — database URLs, Stripe and AI provider keys, AWS role credentials — back into the public repos, roughly every 11 minutes when the analysis was published. If you run Next.js, check request logs for the x-middleware-subrequest header and requests toward the cloud metadata address, and rotate anything exposed.
Affected: React Server Components 19.0.0, 19.1.0, 19.1.1 and 19.2.0 · Disclosed 18 Nov · Under attack: unconfirmed · We had it 18 Nov, 23:06 UTC
Perforce Puppet — Perforce Puppet — CVE-2026-115731, reported at CVSS 9.4, added to CISA's Known Exploited Vulnerabilities catalogue citing active attacks; affected versions and fix per the vendor's advisory. Only a news report is available here — pull CISA's catalogue entry and Perforce's advisory for exact versions and the federal patch deadline.
Holes in AI tools
github.com · today
vLLM
Takeaway: fixed in 0.34.0
VLLM's /v1/systemone endpoint accepts a caller-supplied chat template even on servers started without --trust-request-chat-template, because its code path reaches the shared renderer without passing the validation check other endpoints run. Anyone who can call the API can then control how a conversation is serialised, including replacing the system block an operator pinned, and push in special tokens that per-message checks would block. Defaults are affected — no flag is needed.
It matters on any deployment where API-key holders are not fully trusted and the model is one of the supported architectures, several Qwen3 variants among them. The advisory keeps its claim narrow: a protection-mechanism bypass rated Moderate (CVSS 6.5), with each render still capped at 30 seconds, and no assertion that the Jinja sandbox escapes. Affects 0.32.0rc0 through 0.33.0; fixed in 0.34.0. Reported by Haile Du and Yuan Luo of BAAI, with the report itself filed using AI assistance.
Affected: unconfirmed, see the advisory · Fixed in: 0.34.0 · Disclosed 18 Nov · We had it 18 Nov, 23:05 UTC
rapid7.com · yesterday
Ollama
Rapid7's Christiaan Beek filtered an internet-wide scan of exposed AI services down to 7,675 Ollama daemons answering API calls without authentication. Their model lists double as a record of who has already been there: names carrying the host's own address, models pulled from attacker-chosen IP addresses and throwaway tunnels, and on 55 hosts encoded strings that decode to file paths such as /etc/shadow.
Ollama's API lets any caller pull, copy and push models, so an open port means strangers can confirm write access and outbound connectivity, leave markers, and reach toward the host's files; 79% of these boxes also list cloud-routed models, which can put the owner's ollama.com billing in other people's hands. If anyone on your team rented a GPU server for a quick experiment, this is what the internet sees on it. About 65% of exposed hosts sit in hosting or cloud networks, six providers account for 42% of them, and only 0.3% of findings were served over HTTPS.
“the vulnerability isn't the main problem. The main problem is that nobody owns the server.”
Patches worth knowing
PHP — All supported PHP versions — security releases due 14:00–14:30 UTC today, highest severity critical but non-default setups only; upgrade promptly; pre-set session.upload_progress.enabled=0 for CVE-2026-115560. Fixed version numbers publish alongside the releases; the critical issues need non-default configuration or uncommon setups, per the PHP security team.
Affected: all supported PHP versions · Disclosed 18 Nov · We had it 18 Nov, 23:04 UTC
OpenJDK (Java4Shell) — OpenJDK 17.0.6 and later on AVX-512 VBMI x86-64 — crafted Base64 triggers an out-of-bounds write (Java4Shell); update Corretto or set -XX:-UseBASE64Intrinsics; CVE-2026-114938, Important (CVSS 8.8). Graviton (ARM64) workloads are not affected and AWS says it has no evidence the flaw has been leveraged; restart Java applications after updating — no reboot needed.
Affected: OpenJDK 17.0.6 and later on x86-64 processors that support AVX-512 VBMI · Fixed in: Corretto 17.0.21.1.2, 21.0.13.1.2, 25.0.5.1.2, 27.0.1.1.4 · Disclosed 19 Nov · We had it 19 Nov, 03:31 UTC
AutoGluon MultiModal — Object-detection dataset prep tool before 1.6.0 — downloads dataset archives over plaintext HTTP without validating members, so an on-path attacker can write files outside the target directory; fixed in 1.6.0; CVE-2026-130638. The training and inference APIs, including object detection with MultiModalPredictor, are not affected; if you can't upgrade, skip prepare_detection_dataset and pull datasets over HTTPS instead. AWS rates the bulletin Important.
Affected: before 1.6.0 · Fixed in: 1.6.0 · Disclosed 18 Nov · We had it 19 Nov, 03:31 UTC
One thing to do
Check PHP session settings before today's release lands. If you find any, set session.upload_progress.enabled=0 in php.ini or the FPM pool (not ini_set or .user.ini, which act too late) and restart every PHP worker. aikido.dev
