Ruby, Rails & AI Daily — September 10, 2026

Kicking off daily coverage with a patch-now security story, a JIT performance win, and a busy week for AI in the Rails ecosystem.

Patch now: Active Storage RCE is being actively exploited

Ruby Weekly #815 flags that CVE-2026-66066 ("KindaRails2Shell"), a critical (CVSS 9.5) arbitrary-file-read and remote-code-execution flaw in Active Storage's libvips-based image variant processing, is now seeing active weaponization in the wild. The bug lets an attacker upload a crafted file through Active Storage's direct-upload endpoint and trigger file disclosure via a MATLAB/HDF5 file chain in libvips's "untrusted" operations. It was disclosed back in July with fixes already available — Rails 7.2.3.2, 8.0.5.1, and 8.1.3.1 — but if you're still on an earlier patch level, Rapid7's writeup is worth a read before you do anything else today.

Ruby: ZJIT learns to inline GC allocations

The issue's lead story, written by Peter Zhu, explains that ZJIT can now inline allocations instead of routing every one through a function call — a change that's been on TruffleRuby's side of the fence for a decade and may land in Ruby 4.1. No hard benchmark numbers yet, but it closes a long-standing gap in Ruby's newer JIT pipeline. Also in the issue: Samuel Williams is proposing to drop the "experimental" label from IO::Buffer in 4.1, and RubyGems 4.0.20 shipped support for RUBY_BOX=1.

Rails: chasing Ractors, and Campfire compiled to a 5MB binary

This Week in Rails (Sept 4) covers continued work toward Ractor-ready Rails and improvements to ordered cache fetches. On the more experimental end, Sam Ruby compiled Rails' own Campfire app down to a 5MB binary using Matz's AOT compiler (Spinel) — still incomplete, but a fun signal of where Ruby's ahead-of-time compilation story is headed.

AI meets Rails: Fable 5.1 tops the leaderboard, RubyLLM 2.0 in the wild

The Rails Foundation's Agents on Rails benchmark — which tests frontier and open-weight coding agents against real Rails tasks — now has Claude Fable 5.1 matching Opus 5 at the top with 58/63 successful runs, while costing less and running faster; GLM 5.3 Flash also joined the leaderboard this round. Separately, a new monthly Ruby meetup in Berlin kicked off with a talk on RubyLLM 2.0, which is moving toward a broader provider/protocol model per Carmine Paolino's writeup.

Web development: Chrome 153 ships 230 security fixes

Google's September stable release bumped Chrome to 153.0.8010.36/.37, fixing 230 vulnerabilities across WebGL, Cast, ANGLE, and PDFium — including CVE-2026-87491, which Google says already has an exploit in the wild. If your team auto-updates Chrome fleet-wide, this is a good week to confirm it actually happened.

AI at large: the cybersecurity model arms race

Google, Anthropic, and OpenAI all announced cyber-focused models within days of each other. Google launched Gemini 3.8 Flash Cyber through a new "Fairwind Program" for vetted defenders, partnering with CrowdStrike and Palo Alto Networks. Anthropic paired Claude Fable 5.1 and the more restricted Mythos 5.1 with new "Enterprise Frontier Safeguards," while also disclosing it paused external pre-release evaluations after unauthorized-access incidents. OpenAI said its upcoming Astra model has crossed its internal "Critical" cybersecurity threshold, discovering two zero-days during its own evaluation. Expect more of this framing — CNBC is already calling it "model fatigue" as labs ship at a pace even engineers are struggling to track.

On the radar: Rails World 2026 runs September 23–24 in Austin — worth watching for announcements that build on this week's Ractor and AI-benchmark threads.