Thursday, 1 October 2026
Guardrails Off, For the Trusted Few
Google Shipped Its New Frontier Model With the Cyber Guardrails Removed — For a Chosen Few
Google announced Gemini 4 Argon on Wednesday, pitched as its “next era of frontier intelligence” and built for deep reasoning over long-horizon work: complex coding, enterprise knowledge work and cyber defence. The numbers are good and they are specific. The output ceiling goes to 1 million tokens, up from 64,000. DeepSWE v1.1 comes in at 77.9%, against 74.2% for Claude Opus 5.5 and 74.1% for GPT-6 Astra. On CWE-bench v1, the vulnerability-remediation benchmark, it ties for first at 68%. Introductory pricing is $2 per million input tokens and $10 per million output, settling later to $4 and $20, with a 95% discount on cached input. But the part worth reading twice is the distribution plan. Access runs first through something called the Fairwind Program — US government users and “trusted cyber defenders” — and Google says those testers, along with its own internal teams, get the model without its cyber guardrails, specifically so it can work at full strength finding, validating and patching vulnerabilities. Two days earlier, Sundar Pichai stood in the White House and signed a one-page accord committing his company to monitoring its models for cyber, bio and chemical risk. Both things are true, and they are not necessarily in conflict. They are, however, the same week.
Source: 9to5Google, September 30, 2026
The Pentagon Is Standing Up a Four-Star Command for Autonomous War
Defense Secretary Pete Hegseth used his State of the Force address at Marine Corps Base Quantico on Wednesday to announce AUTOWARCOM — the Autonomous Warfare Command, a new four-star combatant command with “service-like authorities,” built to scale autonomous and robotic capability across the joint force. It still needs Congressional approval, and it does not stand up until October 1, 2027, so what exists right now is an interim effort: Project Agincourt, run out of the Pentagon’s unmanned-systems portfolio office, with Defense Innovation Unit head Owen West as its chief executive and Navy SEAL Senior Chief Max Strasiser as its chief operating officer. The organisational form is the news here, more than the drones. A programme can be cancelled and a task force can be wound up; a combatant command with service-like authorities is the structure the US military reaches for when it expects to be doing something permanently, with its own budget and its own four-star arguing for it.
Source: The War Zone, September 30, 2026
Gartner Expects Most Companies to Walk Away From the Agents Their Vendors Built Them
A forecast worth pinning to the wall for two years: by 2028, Gartner reckons 70% of enterprises will abandon agentic AI built through vendor “forward-deployed engineering” — the arrangement where a vendor’s own engineers embed with the customer and build the thing on site. The reasoning is unglamorous and instantly recognisable to anyone who has ever inherited somebody else’s code. Heavily customised deployments turn out to be expensive to maintain, hard to scale and awkward to govern, and the customer cannot evolve them without the people who wrote them. Gartner also names a practice it calls “FDE washing”: ordinary implementation, professional-services and consulting work relabelled “forward deployed” to justify premium fees without the delivery depth to match. The firm’s advice, from senior director analyst Mukul Saha, is about paperwork rather than technology — “FDE success starts with getting the engagement structure right, from scope and incentives to governance, ownership, and exit.” Four of those five words are things you negotiate before the pilot. The fifth is the one nobody wants to raise in a kickoff meeting.
Source: TechEdge AI, September 30, 2026
DeepMind Put a Watermark Inside AI-Designed Proteins — and It Survives Into the Molecule
Also out of Google on Wednesday, and published in Nature: SynthID Bio, a pair of methods for marking AI-designed proteins so their origin can be traced. For sequences, a SynthID Bio version of ProteinMPNN nudges amino-acid choices as it generates, hiding a signature in the protein’s own code; paired with DeepMind’s AlphaProteo binder design and tested in the wet lab against three real targets — VEGF-A, the SARS-CoV-2 spike receptor-binding domain and PD-L1 — the watermarked designs matched unwatermarked ones on hit rate, binding affinity and sequence diversity. For structures, a fine-tune of part of AlphaFold 3’s diffusion network bakes the mark into the weights, so predicted coordinates carry it regardless of who runs the model. The target user is not a reader but a screener: the companies that synthesise DNA to order, who would get an automatic signal that a sequence came from a model with safeguards attached. DeepMind is refreshingly direct about the limit — resistance to deliberate tampering is not solved, and watermarks need provenance records and central repositories alongside them. The genuinely new thing is the first half: a watermark that stays legible after the design stops being a file and becomes a physical molecule.