Today’s edition: Google ships a mid-priced frontier model, two critical NGINX flaws land with working exploits withheld, and a Microsoft agent package on PyPI turns into a credential stealer. Window covered: roughly 18 November, 06:00 to 19 November, 06:00 (Europe/Rome).
Top 3
- Gemini 4 Neon launches at $3/$15 per million tokens with a 1M-token context window.
- NGINX HTTP/2 RCE pair (CVSS 4.0 9.2 each): patch to 1.30.7 / 1.31.9 or switch HTTP/2 off.
- Microsoft Agent Framework on PyPI was hijacked for two releases that steal cloud credentials at Python startup.
Models and platforms
Gemini 4 Neon (18 Nov). Google positions Neon between Gemini 4 Carbon (deep, long-horizon work) and Hydrogen (fast, everyday execution). List price is $3 per million input tokens ($0.15 cached) and $15 per million output, with a 1M-token window. Google reports 61 on the Artificial Analysis Intelligence Index (third overall), 65.2% on Terminal-Bench 4.1 and 56.3% on Zapier’s AutomationBench. These are vendor-published figures, so wait for independent runs before switching production workloads. It is rolling out in the Gemini API and AI Studio, in Antigravity for paid Pro and Ultra users, and in the Gemini app. A cyber-tuned variant goes only to vetted defenders through Google’s Fairwind Program. Why it matters: if you pay Opus-class prices for agent loops, run your own eval on Neon at the high thinking level. (Google, Antigravity)
Genie 4 (18 Nov). Google’s interactive world model reaches AI Ultra subscribers inside Project Genie; Pro users can join a waitlist. New: you can change the world with a text prompt mid-session, and it generates synchronized audio. The model card lists 720p at up to 24 fps and admits drift in object permanence and legible text. Why it matters: a research preview for now, not something to build on. (Google, model card)
OpenAI on chip design (18 Nov). OpenAI says its models found up to 14% power/performance/area gains over heavily optimized baselines on AMD flagship GPU IP, and cut standard-cell area by about 13% (vector unit) and 10% (matrix unit) in its own JalapeƱo inference chip. It also announced a Synopsys partnership to train a model that drives EDA tools. All numbers are OpenAI’s own and the RISC-V evaluation has acknowledged simplifications. Why it matters: a concrete example of agents doing verified work in a toolchain, with a human-checked loop. (OpenAI)
Claude and Anthropic. No new announcement appeared on anthropic.com/news or the claude.dev blog inside the window; the latest items there date from 9 and 17 November. Claude Code 2.1.328 shipped on 19 November, and the only change I saw in its notes is a VS Code notice when a message from another session is held. Separately, Anthropic is credited, with Ada Logics, for four of the seven Ruby bugs below. Why it matters: quiet day for Claude releases; read the 2.1.328 notes before updating if you run VS Code sessions. (Claude Code releases)
Security
NGINX: CVE-2026-123633 and CVE-2026-123007 (18 Nov). Both are unauthenticated remote code execution bugs in NGINX’s HTTP/2 handling, rated 9.2 on CVSS 4.0. One is a use-after-free that matters mainly when NGINX proxies and runs low on connection slots; the other is a flow-control integer underflow that Wiz exploited on a server serving plain static files over TLS, with an ASLR bypass. The latter reaches back to version 1.9.5 (2015). Fixed in 1.30.7 (stable) and 1.31.9 (mainline). If you cannot upgrade, set http2 off; (NGINX 1.25.1 or later) or drop the http2 parameter from listen, then reload. No in-the-wild exploitation was reported and neither flaw was in CISA’s KEV catalog as of 18 November; Wiz is withholding its exploits. Why it matters: patch internet-facing NGINX first, and check when your distro backports land. (Wiz, The Hacker News)
Microsoft Agent Framework on PyPI (18 Nov). Versions 1.26.1 and 1.26.2, published on 17 November, carried a stealer in .pth files that runs at Python startup, even if the package is never imported. It targets Azure, AWS, GitHub, Docker, Hugging Face and OpenAI credentials, plus Azure managed-identity and CI OIDC tokens, and exfiltrates to a webhook.site URL. PyPI quarantined both; how the attacker got publish rights is unknown. The package sees roughly 200,000 downloads a month. Why it matters: if either version was ever installed, rebuild from a clean environment on 1.26.0, rotate every reachable credential and search egress logs for webhook.site. (BleepingComputer)
Ruby: seven memory-safety bugs (19 Nov). The Ruby team fixed seven flaws in the regular-expression engine and string case conversion, and says it knows of proofs of concept that reach RCE for several. CVE IDs are not assigned yet. Upgrade to 4.0.10, 3.4.13 or 3.3.14; Ruby 3.2 and earlier get no fix. Why it matters: highest risk is any app that compiles user-supplied regexes or case-converts untrusted strings. (ruby-lang.org)
Linux KVM: CVE-2026-99312 (18 Nov). A stale MMU context when emulating INIT can crash or corrupt a host, but only on AMD hosts with nested virtualization. Red Hat rates it moderate (CVSS 5.6); fixed in 6.12.118, 6.18.64 and 7.3.5. Why it matters: only urgent for hypervisor hosts that offer nested virtualization to untrusted guests. (linux-cve-announce, Red Hat)
Open source and local AI
OVH Olala mini 7A1B. OVH announced its first in-house LLM family on 17 November, with the official launch at its AI Summit on 19 November. The model is a 7B mixture-of-experts with 1B active parameters and a 32K context. The company says it runs on an 8 GB GPU and at about 50 tokens per second on a plain CPU. Weights are due on Hugging Face under Apache 2.0, with base model, technical report and AI Act documentation. I could not confirm that the repository is live yet, and the speed figures are OVH’s. Why it matters: a small, permissively licensed European model to benchmark for CPU-only or edge deployments. (OVH)
Sources
- Google: Introducing Gemini 4 Neon
- Google Antigravity: Gemini 4 Neon
- Google: Genie 4 is now available
- Genie 4 model card
- OpenAI: Advancing chip design with OpenAI models
- Claude Code releases (GitHub)
- Wiz: NGINX HTTP/2 RCE vulnerabilities
- The Hacker News: NGINX HTTP/2 flaw
- BleepingComputer: Agent Framework PyPI compromise
- Ruby security advisory
- linux-cve-announce: CVE-2026-99312
- Red Hat: CVE-2026-99312
- OVH: Olala mini 7A1B
