Agents expand into office, finance; regulatory hurdles grow

Anthropic / Claude ecosystem

No significant new developments.

Frontier model providers

OpenAI releases major Codex update for office tasks - YouTube

OpenAI has expanded Codex from a coding-only tool to a general-purpose desktop AI agent for office tasks, featuring a dynamic role-based UI and integrations with Slack and Google Workspace. This allows Codex to function as a personal work assistant for teams, automating non-technical productivity tasks.

OpenAI Expands Advanced Voice Mode Access to Web Users Without Direct Video Support

OpenAI is expanding its Advanced Voice Mode to desktop browsers, providing real-time conversational AI capabilities for web users. This enables creators and professionals to utilize advanced voice interactions without requiring video support.

Google DeepMind teases RT-3 with new capabilities - RobotWale News

Google DeepMind has teased its RT-3 robotic foundation model, demonstrating improved generalization and dexterity for humanoid robots. This advancement enhances long-horizon planning and natural language understanding, pushing vision-language-action modeling for general-purpose robotics.

Meta Introduces Autodata: An Agentic Framework That Turns AI Models into Autonomous Data Scientists for Building High-Quality Training Data - Dataforcee Digital

Meta's Autodata framework uses AI agents as autonomous data scientists to iteratively generate and refine high-quality training data. This agentic self-instruct implementation significantly outperforms traditional CoT Self-Instruct, widening the performance gap between weak and strong solvers from 1.9 to 34 percentage points.

DeepSeek Finally "Opens Its Eyes": Multimodal Image Recognition Goes Live, the Last Missing Piece for Chinese LLMs - DEV Community

DeepSeek has officially added multimodal image recognition capabilities with visual-reasoning fusion to its models. This integration completes its feature parity with global competitors like GPT-5.5 and Claude 4.

Google releases experimental ‘COSMO’ AI assistant on Play Store

Google accidentally released an experimental AI assistant app named 'COSMO' on the Play Store, which features local Gemini Nano model capabilities. The app was subsequently removed shortly after its release.

Google Unveils Powerful AI Agent, Boosts Cloud Capabilities – OVEX NEWS

Google has launched the Gemini Deep Research agent, which reportedly outperforms benchmarks and accelerates scientific analysis from weeks to days. Additionally, Google unveiled CX Agent Studio, designed for enterprise customer service automation.

AI developer tooling & infrastructure

GameMaker Adds Claude Code to GM-CLI for AI-Assisted Game Development

GameMaker has integrated Anthropic's Claude Code as the first native AI-assisted developer agent into its command-line toolchain (GM-CLI). This enables game developers to use natural language for project queries, debugging, and build management directly outside the IDE.

The AI scaffolding layer is collapsing. LlamaIndex's CEO explains what survives. | VentureBeat

The CEO of LlamaIndex argues that the 'AI scaffolding layer' is collapsing as frontier models gain advanced reasoning capabilities. This shifts the competitive advantage from complex frameworks to effective context extraction and document parsing.

JobRunr Introduces ClawRunr, an Open-Source Java AI Agent - InfoQ

JobRunr has launched ClawRunr, an open-source Java AI agent designed for conversational interaction and persistent task execution. It addresses challenges like retries, persistence, scheduling, and monitoring for long-running agent workflows.

Cloud & platform providers

No significant new developments.

AI policy, regulation & governance

OpenAI is Now Sharing Its Users' Data With Advertisers

OpenAI's updated privacy policy formally states that it will share user data with advertisers and receive purchase data from marketing partners. This change signals OpenAI's strategic shift toward an ad-supported platform model.

CAISI Evaluation of DeepSeek V4 Pro | NIST

NIST's independent evaluation of DeepSeek V4 Pro indicates an approximately 8-month capability lag compared to U.S. frontier models. The evaluation also found lower performance on agentic reasoning benchmarks than DeepSeek's self-reported claims.

Colorado’s AI compromise would focus regulations on informing consumers when the technology is used - The Colorado Sun

Colorado has introduced a third compromise attempt to rewrite its AI regulation law, Senate Bill 189. This new bill shifts the focus from developer disclosure to consumer notification requirements and delays implementation until January 2027.

US Senate Judiciary tees up AI chatbot, companion safety debate | IAPP

The US Senate Judiciary Committee is advancing a bipartisan bill, the Guidelines for User Age-verification and Responsible Dialogue Act (GUARD Act), aimed at protecting children from harmful AI chatbot content. The bill proposes age verification and disclosure requirements for AI chatbot providers.

Senators take another swing at bill to codify federal AI resource | FedScoop

The U.S. Senate has reintroduced the bipartisan CREATE AI Act (S.4441), aiming to codify the National Artificial Intelligence Research Resource (NAIRR). This initiative seeks to democratize access to essential AI research tools and computational resources.

Senate Panel Backs GUARD Act, AI Age Verification Bill

The U.S. Senate Judiciary Committee has unanimously advanced the GUARD Act, a bill that would require AI chatbot providers to verify user age through government ID, facial scans, or financial records before granting access.

The Colorado AI Act Hits a Wall: Litigation, Legislative Uncertainty, and an Enforcement Standstill | Privacy World

A federal magistrate judge has issued a stay of enforcement for Colorado's AI Act (SB24-205) just weeks before its effective date. This action halts state enforcement and creates legislative uncertainty regarding a potential replacement bill.

Clean up AI or go away: rules proposed for data centres - Yahoo News Australia

A NSW inquiry proposes new conditions for data centre development approvals, mandating guarantees that hosted AI models do not violate copyright law and utilize renewable energy. This aims to regulate the environmental and legal impact of AI infrastructure.

A Chinese court has ruled that replacing an employee with AI alone does not constitute legal grounds for termination without meeting other established legal dismissal criteria. This ruling stems from a case involving a quality assurance supervisor in Hangzhou.

Oscars bans AI actors and writing from awards

The Academy of Motion Picture Arts and Sciences has formally amended its eligibility rules, explicitly banning AI-generated acting and writing from Oscar consideration. This marks the first direct restriction on generative AI in the awards' history.

Industry & market moves

OpenAI starts laying foundations for ChatGPT ads in EU

OpenAI is rolling out a GDPR-compliant conversion tracking pixel in the EU, laying the groundwork for ChatGPT advertisements. This marks the initial infrastructure deployment for OpenAI's advertising expansion beyond the US market.

Meta buys Assured Robot Intelligence to boost humanoid robots, expands Meta Superintelligence Labs and Meta Robotics Studio - CNBC TV18

Meta has acquired robotics AI startup Assured Robot Intelligence (ARI) to accelerate its humanoid robot development. ARI co-founders Lerrel Pinto and Xiaolong Wang will join Meta Superintelligence Labs, bolstering Meta's efforts in this field.

Nebius acquires AI model optimization startup Eigen AI for $643M - SiliconANGLE

Nebius, a Dutch AI data center operator, is acquiring Eigen AI, a model optimization startup, for $643 million. This acquisition aims to enhance Nebius's Token Factory managed inference service by boosting AI inference token optimization on Nvidia chips.

Encord Announces $60 Million in Series C Funding

Encord has raised $60 million in Series C funding to scale its AI-native data infrastructure, specifically for physical AI systems moving towards production deployment. This funding aims to support the growing demands of industrial AI applications.

AI product & feature launches

Zensai Introduces Human Success Agent for Microsoft Agent 365, Redefining AI-Driven Learning and Human Success in the Flow of Work

Zensai has launched the Human Success Agent for Microsoft Agent 365, enabling enterprises to embed AI-driven workforce insights directly into Microsoft 365. This offers secure governance through Microsoft Entra for AI-driven learning and human success initiatives.

Microsoft wants lawyers to trust its new AI agent in Word documents | The Verge

Microsoft is launching a specialized AI agent for legal teams within Word, designed to review contracts and identify risks. This new tool leverages expertise from engineers of the recently acquired Robin AI.

Citi launches Arc to scale AI agents across the business | CIO Dive

Citigroup has deployed 'Arc,' an enterprise-scale agentic AI platform aimed at automating manual tasks such as research and data analysis across all its business lines and global geographies.

Kamiwaza AI Launches Kamiwaza 1.0: Secure AI Orchestration Built for Regulated Industries

Kamiwaza AI has launched Kamiwaza 1.0, introducing governed collaboration spaces (Workrooms), hardened container infrastructure, and an upgraded AI agent. This platform is specifically designed for regulated industries that need to operate across distributed data without centralization.

MoonPay Announces 'MoonAgents Card' Enabling AI Agents to Spend Stablecoins Anywhere Mastercard Is Accepted

MoonPay has announced the 'MoonAgents Card,' which allows AI agents to spend stablecoins directly from on-chain wallets at any merchant globally that accepts Mastercard. This combines crypto-to-fiat conversion with programmatic agent authorization.

Research with immediate practical relevance

A New NVIDIA Research Shows Speculative Decoding in NeMo RL Achieves 1.8× Rollout Generation Speedup at 8B and Projects 2.5× End-to-End Speedup at 235B - MarkTechPost

NVIDIA research demonstrates the integration of lossless speculative decoding into reinforcement learning (RL) training loops, achieving a 1.8x rollout generation speedup at 8 billion parameters and projecting 2.5x end-to-end gains at 235 billion parameters.