AI.info

The Pulse — Page 18

Browse The Pulse on AI.info.

• Tess-4.5 Matches Qwen3.8 With 25% Fewer Tokens

  Continuity Research says Tess-4.5 matched its base model on the company’s BenchLocal evaluation while using 25% fewer tokens across 150 scenarios. The October 2026 release describes the 27-billion-parameter open-weights model’s training data, local speed tests and evaluation conditions.

• NVIDIA Ships Isaac ROS 5.0 With Agent Skills for Robot Developers

  NVIDIA has released Isaac ROS 5.0, adding AI-agent workflows, ROS 2 Lyrical support and new tools for building physical AI applications. The open-source release runs across Jetson Orin Nano and Jetson Thor platforms and includes faster FoundationPose inference.

• Google Rolls Out Gemini Omni 1.2 Flash With Native 15-Second Video Generation

  Gemini Omni 1.2 Flash offers native video generation and editing for clips up to 15 seconds, Google says. Google cut the price by 30% from Gemini Omni 1.1 and added multi-turn editing and, it says, stronger reference following.

• DOJ Seeks Information From Vendors for 20- to 30-User Legal AI Pilot

  Nextgov/FCW reported that the Justice Department is asking vendors about commercially available legal AI platforms for a potential pilot involving 20 to 30 government users.

• Nvidia Ties Southeast Asian Cloud Partner Status to Chip-Tracking Enrolment, Bloomberg Reports

  Nvidia now requires customers in Southeast Asia to enrol in its fleet intelligence system to become official Nvidia Cloud Partners, or NCPs, according to people familiar with the matter cited by Bloomberg. The requirement links partner status to a voluntary tracking programme that tracks the location of participating companies’ AI accelerators, among other metrics.

• Google Expands CC Into a Shared AI Agent for Families

  Google is expanding its experimental CC agent to help families coordinate schedules, tasks and household logistics with permission-based sharing.

• OpenAI’s Asia-Pacific Policy Chief Is Leaving After Six Months

  OpenAI’s Singapore-based Asia-Pacific public policy chief, Sanghyun Lee, is leaving after six months in the role, Bloomberg News reported, citing a person familiar with the matter. OpenAI declined to comment; the company faces regulatory scrutiny in the region.

• FLEX lets robots generalize human corrections

  MIT and New York University researchers introduced FLEX, a system that lets robots use conversational corrections to infer broader preferences. In simulations, it generalized warnings about specific objects to categories such as electronics and fragile items.

• Shontel Brown’s Election AI Bill Would Require Labels

  Rep. Shontel Brown introduced the Securing Elections From AI Deception Act, which would require disclaimers on certain AI-generated election content. The bill also proposes FTC enforcement and penalties for using AI to intentionally interfere with voting rights.

• Watney Raises $80 Million to Put Robots Inside Data Centers

  Watney has raised $80 million in Series A funding to expand robots that perform physical maintenance inside data centers. The San Francisco startup says its systems have logged hundreds of thousands of customer-site hours with reliability above 99.99%.

• OpenAI Delays GPT-6.1 Astra After Safety Review Flags Risks

  OpenAI said it is delaying GPT-6.1 Astra after researchers raised security concerns, with the company’s head of safety systems saying the model did not meet its bar. The decision comes as OpenAI describes a new process for reporting misalignment and reviews agents’ unexpected activity on government websites.

• Ridge Security Launches RidgeOne, a Local Red-Team Appliance

  Ridge Security launched RidgeOne, an AI penetration-testing appliance built around NVIDIA DGX Spark that runs an open-weight model and RidgeGen locally. The company says a test of OWASP Juice Shop found 32 verified issues in about 10 hours, while deployment requires hardware and a RidgeGen license.

• Anthropic Sets Up a 2028 Presidential Engagement Program

  Anthropic is hiring a Political Programs Lead to build a presidential engagement program for the 2028 election cycle, with a listed salary range of $295,000 to $345,000. The role combines candidate briefings on AI policy with political strategy and oversight of the company’s political giving programs.

• Mistral Agents Move 40,000 Fortran Lines to C++

  Mistral AI says it helped a European energy operator migrate 40,000 lines of Fortran 77 from a reservoir simulator to C++. The company’s case study shows that testing, documentation and human review mattered more than giving coding agents full autonomy.

• Researchers Use AI to Find 564 New Magic-State Factories

  A new arXiv paper reports 564 previously unreported magic-state factory classes found through symmetry-guided searches and language-model agents. The catalog includes protocols for producing T, CS and CCZ states, but the authors say further work is needed to measure their costs in a full quantum-computing architecture.

• Lambda Cloud Adds Free Role-Based Access Control

  Lambda Cloud's identity and access management brings role-based access control to Lambda Cloud, with roles, groups and service identities in the console. IAM is free, and audit logs are kept for six months.

• OpenAI’s GPT-6 Astra Scores 87.4% Across 20 Coding Languages

  On MultiLangBench, which measures software-building across 20 programming languages, GPT-6 Astra achieved an 87.4% average test pass rate, OpenAI says. COBOL migration results also improved, the company reports, while it warns that production rewrites still need expert review and testing.

• NVIDIA details 2B-parameter model for onboard AV reasoning

  NVIDIA introduced Alpamayo 2 Edge, a vision-language-action model designed for real-time use on its DRIVE AGX Thor platform. The company reports a 160–190 ms pixel-to-trajectory latency for an optimized configuration and describes AlpaJudge workflows for evaluating and improving model reasoning.

• Claude Desktop Flaw Ran Malicious Git Executables in Repositories

  Anthropic disclosed a Windows flaw that could run an attacker’s Git executable when a user opened a project in Claude Desktop. Version 1.8089.0 fixes the vulnerability, which the advisory rates High with a CVSS v4 score of 7.5.

• Xiaomi Open-Sources MiMo-V2.6 and Its Reinforcement-Learning Stack

  Xiaomi announced MiMo-V2.6 on September 22, releasing Pro and Flash model weights alongside reinforcement-learning code, training environments and a technical report. The company’s six-day training run completed about 750,000 trajectories, while its published benchmark results show strengths and gaps against leading closed models.

• GPT-6 Astra Edits Qwen Weights, Raising Math Score to 48.1%

  MIT CSAIL researchers say GPT-6 Astra edited 0.016% of Qwen3.5-0.8B’s weights without gradient training. On MATH-500, the edited model scored 48.1%, up from 7.9%, in the team’s evaluation.

• Giotto Launches Alto and Solo Sovereign Reasoning Models

  Swiss AI lab Giotto has launched Alto and Solo, two reasoning models designed for European cloud, dedicated hosting and on-premises use. Alto runs on eight B300 GPUs, while Solo is designed for a single GPU.

• Microsoft brings Copilot and agent governance to GCC on October 1

  Microsoft is launching Microsoft 365 G7 and Agent 365 for Government Community Cloud customers on October 1, 2026. The package combines Copilot, mission-focused agents, identity, security, compliance and controls for managing AI deployments.

• Aachen Court Extends Google AI Search Liability Ruling

  Aachen Regional Court barred Google from repeating false claims generated by its AI Mode search feature about a doctor. The September injunction adopts reasoning from a Munich case, where a separate ruling against Google later became moot after a settlement.

• Microsoft Gives Copilot Studio Agents Their Own Accounts

  Microsoft is adding agent accounts and business-process memory to Copilot Studio so organization-managed agents can work across Microsoft 365. The features are rolling out through the Frontier program, while additional administrative controls are planned for later.

• OpenAI Adds Cathedral to Security Sprint and Raises Bug Bounties

  OpenAI plans to bring Cathedral researchers into its next internal security sprint and has raised rewards for its Security Bug Bounty program. The company says exceptional critical findings on core products may qualify for up to $172,000 with bonuses.

• Meta Announces Enterprise AI Platform Led by CJ Desai

  Meta announced an enterprise platform that will bring its AI models, agents, infrastructure and developer tools to businesses and developers. Chirantan “CJ” Desai will lead the effort as chief enterprise platform officer, reporting to Mark Zuckerberg.

• Reliable Robotics Raises $115 Million as FAA Certification Advances

  Reliable Robotics raised $115 million, lifting its Series D to $275 million and its valuation above $1 billion. The company plans to expand its aircraft fleet, production facilities and engineering team as it pursues FAA certification for automated flight.

• Nine AI Agents Produce Zero Passing Scientific Follow-Ups

  The Into the Unknown benchmark tested nine AI agents on 15 open research problems contributed by scientists affiliated with U.S. Department of Energy national laboratories. None of 135 submissions met every condition, and 114 failed integrity checks for claims that did not match the work or its recorded evidence.

• INFUSE Finds 19.6% of Agent Attack Trials Were Compromised

  A benchmark called INFUSE tested indirect prompt injection across 40 agent tasks and 19 language models. The authors report a 19.6% mean attack success rate under an LLM judge, while warning that measured defenses and results depend on the test setup.

• Walmart Opens Its Shopping Systems to AI Agents

  Walmart announced Agent Profiles and a new Agent Gateway for AI agents shopping at Walmart and Sam’s Club. Customers can set purchase approvals and spending limits, while agents cannot access payment-card details.

• BNP Paribas Signs Five-Year Google Cloud AI Partnership

  BNP Paribas and Google Cloud have agreed to a five-year partnership covering cloud infrastructure, Gemini models and Gemini Enterprise. Planned first deployments include AI agents for corporate credit memos and Gemini integration in an internal assistant available to more than 65,000 employees.

• GPT-5 Help Cut Persistence on Follow-Up Tasks

  Three randomized experiments involving 1,222 people found that brief access to GPT-5 improved performance while available but was followed by weaker results on unaided tasks. The tests measured responses immediately after assistance ended, and one replication did not find a significant overall increase in skipping.

• Epoch AI Switches Capabilities Index to a Bayesian Model With Better-Justified Uncertainty Estimates

  Epoch AI has updated the modeling behind its Epoch Capabilities Index with a Bayesian model. The model provides better-justified uncertainty estimates and lets Epoch AI calculate how likely one model is to be more capable than another.

• AI Labs Game Out the Politics of a Catastrophic Attack

  Executives at Anthropic, OpenAI and other AI companies are privately planning for political and public backlash after a catastrophic AI-related event. Axios reports that industry insiders expect such an event within six to 12 months, while Congress debates measures including an AI kill switch.

• AI Frontiers Is Over After Its Early-September Reorganization Into Microsoft AI, John Langford Writes

  AI Frontiers is over after it was reorganized into Microsoft AI in early September, John Langford says. He describes the group’s research projects and writes that there will be no internships with the people there in the coming year.

• BCG says industrial AI leaders grow revenues more than twice as fast as laggards

  BCG says industrial companies leading in AI grew revenue more than twice as fast as sector laggards from 2022 to 2025. Its modeling estimates that fully AI-first companies in three industrial subsectors could raise EBIT margins by 2.5 to 5.6 percentage points at maturity.

• Sierra Rebuilds Agent Studio Around Ghostwriter

  Sierra has redesigned Agent Studio around Ghostwriter, bringing investigation, journey updates and testing into one conversation. The update also groups related agents into projects and adds review and release controls for changes.

• Santander and Nvidia Set Three-Year AI Plan for the Bank

  Santander and Nvidia have agreed to a three-year collaboration covering banking AI, including models, agents, fraud prevention and simulation. Early experiments reported by the companies suggest smaller, specialized models can reduce serving costs, while Santander’s separate AI target is more than €1 billion in total business value over 2026–2028.

• AlphaPai Raises $50 Million for Financial AI Research Platform

  AlphaPai has raised US$50 million in Series B funding to expand its financial AI research products and reach more institutional investors. AlphaPai will direct the capital toward model development, multi-asset coverage, industry partnerships and overseas growth.