The Signal — October 15, 2026

Google's most capable model is now available to paying developers, two Nature papers show AI design tools producing enzymes that cut chosen proteins, and Figma wants teams to write down the rules their design agents should follow.

Gemini 4 Argon is now open to paying developers and subscribers, two weeks after a restricted debut

When Google announced Gemini 4 Argon on September 30, it gave the model first to vetted cyber defenders, which we covered in our October 1 edition. On October 13 Google opened it up: developers and enterprises can now use Argon through the Gemini API, Google AI Ultra subscribers can use it in the Antigravity coding tool, and Ultra and Pro subscribers get it in the Gemini app and AI Mode in Search. The API costs $2 per million input tokens and $10 per million output tokens until November 30, then doubles to $4 and $20. GitHub added Argon to Copilot the same day for Pro+, Max, Business and Enterprise users, with a gradual rollout.

Google is also widening Fairwind, the program that gives approved security teams a version of Argon without its cyber guardrails, to "hundreds of additional" partners. In its launch post, Google reports an increase of more than 35% in agent-written code submissions inside the company, with nearly a third fewer rollbacks, and a Chromium refactor finished in weeks instead of months. Its safety results include external biology red-teamers unlocking under 1% of the tested sub-tasks and a 0.7% success rate for attackers on Gray Swan's indirect prompt-injection benchmark. These are Google's figures, not independent measurements.

Not every Gemini user gets it. The Decoder notes that free and Plus subscribers are limited to the smaller Flash models under Google's new plan structure. By the outlet's reading of independent overall rankings, Argon still sits behind Anthropic's Claude Opus 5.5. The Decoder also says Google's write-up does not make clear whether Argon's prompt-injection score came from the same test on which rival models were scored earlier.

Sources: Google · The Decoder · GitHub Changelog

AI design tools produced enzymes that cut chosen proteins, including ones linked to Alzheimer's

Proteases are the enzymes that break the peptide bonds holding proteins together, bonds that would otherwise last for centuries in water. Two papers published in Nature on October 14, primarily from David Baker's lab at the University of Washington, describe proteases designed from scratch with AI protein-design tools and aimed at specific target sequences. As Ars Technica explains, the team kept the handful of catalytic amino acids from a known enzyme, had RFdiffusion build a new protein backbone around them and the target, filled in the sequence with ProteinMPNN, and checked predicted structures with AlphaFold3.

In the cysteine-protease paper, 13 of 69 designs worked, and six crystal structures closely matched the designs. The designed proteases have folds very different from those of natural proteases, and they cut their targets inside mammalian cells. In the zinc metalloprotease paper, 36% of 135 designs cut their target peptide at the intended bond. Four amino-acid substitutions pushed the best one past a ten-billion-fold speed-up over the uncatalysed reaction. The team then designed metalloproteases that selectively cut TDP-43, amyloid-β and serum amyloid A, proteins associated with neurological and other diseases.

These are lab results, and the designs started out far less active than the natural enzymes they were modelled on. Ars notes that turning them into therapies would mean getting proteins to the right place without setting off the immune system, a hard problem of its own. It also reports a limit the researchers point to themselves: the tools are good at producing a requested structure but cannot yet be asked for a given level of catalytic activity, so much of the later improvement came from swapping amino acids near the active site and testing the variants. The full papers are paywalled; the figures here come from their abstracts and Ars's reporting.

Sources: Nature (cysteine proteases) · Nature (metalloproteases) · Ars Technica

Figma gives teams a place to write the rules its design agents should follow

At its first Config event in India on October 14, Figma launched Specs, a Markdown document that sits on the Figma canvas and can point to designs, FigJam diagrams and working code. Teams can use a spec as shared context for colleagues and for AI agents, and pass it to a coding agent through Figma's MCP server. Specs is rolling out in open beta on paid plans. Figma also added guidelines to design libraries (open beta) and code sources imported from GitHub or npm (closed beta), so that its agent can follow a team's existing patterns. Code layers, which put working code on the canvas, moved to open beta on all plans.

TechCrunch describes the release as Figma's attempt to steer agents whose output varies from run to run toward more predictable results, a week after Figma widened access to its on-canvas agent. Figma's chief design officer, Loredana Crisan, told TechCrunch that 17 times more code is being written than before but releases are up only 1.3 times, and that much of the extra code is written to work out what the product should be. Those figures are Figma's, and the company has not shown evidence yet that written guidelines make agents stick to a brand's rules reliably. The practical change is that the instructions a team gives its agents now sit next to the designs, where colleagues can read and edit them.

Sources: Figma · Figma (Specs) · TechCrunch

On the Editor's Desk

We held NVIDIA's pledge of compute access valued at $250 million over five years for UK agent-safety research, because it overlapped heavily with the UK AI-safety stories in our last two editions and its providers and terms are still to be agreed. POLITICO's report that the EU's cyber agency and Joint Research Centre are testing Chinese open-weight models is from last week, with nothing new since. Microsoft's machine-learning method for two-to-six-week weather forecasts also waits, since we could read only Microsoft's own account of the Nature paper.