Google accelerates Gemini 4, top labs form AI safety body & more

Google DeepMind has accelerated the development timeline for its next-generation foundational model, Gemini 4, advancing the system into its early post-training phase ahead of internal schedules. The Information reported from the AI Agenda Live Summit on September 24, 2026, that Google DeepMind chief Koray Kavukcuoglu confirmed the milestone, stating that the company aims to roll out the flagship model much earlier than the year-end target previously anticipated by industry observers. The summit marked Kavukcuoglu’s first public appearance in the top DeepMind leadership role. Addressing attendees, he emphasized that initial training iterations yielded stronger benchmark signals than projected, prompting the team to compress the timeline. Our intention is to roll out an early post-training version as soon as possible, because we've already seen promising results and are very excited, Kavukcuoglu said. While framed as a declaration of intent rather than an unalterable release date, the announcement signaled a competitive push to preempt rival frontier releases. The post-training stage for Gemini 4 will encompass reinforcement learning from human feedback, safety alignment, agentic fine-tuning, and extended tool-use optimization. The accelerated timeline follows Google's rollout of specialized auxiliary systems, including Gemini 3.8 Flash TTS and Flash-Lite TTS speech models supporting over 100 languages and dialects, which DeepMind has used to test multimodal integration pipelines. By expediting post-training, DeepMind is positioning Gemini 4 to challenge upcoming frontier iterations from OpenAI and Anthropic as the race for complex multi-step reasoning capabilities intensifies.

Anthropic, Google, and OpenAI are coordinating to establish an independent, self-regulatory standards organization for frontier artificial intelligence by late 2026 or early 2027, according to a report published by The Information on September 24, 2026. The initiative is designed to institute shared testing benchmarks, formalize incident disclosure channels, and set voluntary safety and cybersecurity guardrails across frontier developers without relying on government oversight. The push for an industry-governed standards body materialized after earlier attempts to construct a formal public-private evaluation partnership stalled under the Trump administration. Under the current framework, the proposed organization would define qualification criteria for independent model auditors, set uniform evaluation protocols for pre-deployment safety assessments, and establish explicit requirements for reporting unauthorized web access or cybersecurity anomalies. Members of an active working group are also deliberating whether the standards body should directly execute red-teaming and safety testing on frontier foundation models. The initiative interfaces with existing cross-lab channels, including the Frontier Model Forum, a nonprofit founded in 2023 by OpenAI, Anthropic, Google, and Microsoft that remains active and in regular communication with the organizers. High-level momentum for standardized risk controls was echoed on September 24 during a United Nations Security Council meeting, where OpenAI CEO Sam Altman and Anthropic CEO Dario Amodei both testified on the necessity of international AI safety baselines. Concurrently, Microsoft President Brad Smith told Bloomberg Television that Microsoft supports independent model evaluation and advocates for mandatory human-controlled off switches for advanced autonomous systems. However, the proposal has drawn criticism from open-source advocates and independent developers, who caution that a self-regulatory body led by the three largest frontier labs could establish compliance barriers that entrench market incumbents.

Cybersecurity pioneer Island raised $400 million in a Series F funding round on September 24, 2026, lifting the company’s post-money valuation to $6.4 billion. The financing was led by Evolution Equity Partners, with participation from returning and new institutional investors including Prysm Capital, Sequoia Capital, Coatue Management, Cyberstarts, Insight Partners, J.P. Morgan Growth Equity Partners, Alta Park Capital, Georgian, G Squared, and Squarepoint. Dmitri Alperovitch, co-founder and former chief technology officer of CrowdStrike, also made a follow-on personal investment. The transaction more than doubles Island’s valuation since 2024 and follows a $250 million Series E round completed in March 2025 at a $4.8 billion valuation.

Founded in 2020 by CEO Mike Fey and CTO Dan Amiga, the Dallas-based firm is pivoting from its initial identity as an enterprise browser provider into what it terms an agentic control plane. As corporate enterprises increasingly deploy autonomous software agents to query databases and trigger transactions, Island is extending its policy engine and audit logging across browser sessions, endpoint devices, applications, modern Secure Access Service Edge networks, and underlying data repositories. Island CTO Dan Amiga explained that because AI agents operate across multiple layers of the technology stack, enterprises cannot govern them from a single silo. The company's Enterprise AI module introduces identity governance, guardrails, human-in-the-loop approvals, and compute cost boundaries regardless of which foundation models an enterprise adopts.

The massive capital injection reflects rapid top-line growth and deep enterprise penetration. Island has doubled its annual recurring revenue every fiscal year since its commercial launch in 2022, reaching approximately $200 million in ARR while doubling its global headcount over the past year to 1,000 employees. The company disclosed that eight of the world's ten largest banking institutions now rely on its architecture for digital governance.

The round underscores a broader venture surge into the security and identity governance of AI agents. Evolution Equity Partners founder Richard Seewald stated that enterprises are fundamentally rethinking the architecture of workplace computing, positioning Island to govern how humans and machine agents interact with proprietary data. Island faces a competitive landscape that includes legacy network providers like Palo Alto Networks, which acquired browser security startup Talon in 2023, as well as specialized startups like Neo, Cloudflare's agent browser initiatives, and data security platform Cyera, which raised $400 million concurrently.

Feather Robotics emerged from stealth on September 25, 2026, unveiling a $29,990 modular, two-armed humanoid development platform alongside a $7.6 million pre-seed round. The funding was led by Gradient Ventures, with backing from BuilderVC, Geometry, SEED Innovations, and Virgo VC. Unlike vertically integrated robotics developers such as Tesla, Figure, 1X, and Agibot that build proprietary robot bodies and intelligence models concurrently, Feather focuses purely on hardware, marketing itself as an open physical foundation for developers running third-party models from Nvidia, Skild AI, or Physical Intelligence.

The company was founded by Hoa Mai, who previously sold a humanoid startup to 1X, and Parsa Bakhtiari, a former Tesla Model 3 engineer. Gradient general partner Darian Shirazi noted that Feather has already surpassed $1 million in commercial revenue after shipping hardware for roughly a year, while expending only a fraction of its initial pre-seed capital. The founders are positioning the system as the Android of robotics, offering standardized physical hardware to support the anticipated emergence of thousands of specialized physical AI software companies over the next five years.

The platform utilizes a wheeled base rather than bipedal legs, prioritizing mechanical reliability, part reduction, and battery life over all-terrain walking. The 105-kilogram robot features 23 degrees of freedom, including seven per arm, a one-meter arm reach, a 600-millimeter vertical torso lift, and dual hot-swappable 48-volt batteries providing approximately 10 hours of runtime. Its compute bay accommodates customer-swappable Nvidia Jetson processors, ranging from Nano to Orin AGX and Thor, paired with Zed stereo RGB-D head cameras, field-swappable grippers, and power lines designed for integrators. While rated payload at full extension is limited to 1.8 kilograms with a 6.8 kg peak, early units are already deployed in commercial pilot environments, including commercial kitchens and scientific laboratories.

Feather’s launch comes amid significant regulatory and economic shifts in the robotics sector. Following the Federal Communications Commission’s July 2026 import ban on newly certified foreign-manufactured humanoids and quadrupeds, domestic robotics assemblers face reduced direct competition from low-cost Chinese hardware platforms. By focusing on modular domestic assembly, ROS2 compatibility, and an Apache 2.0 open-source Python SDK, Feather is aiming to bridge the cost gap between fragile sub-$5,000 research kits and six-figure industrial humanoid systems.