AI.info

The Pulse — Page 17

Browse The Pulse on AI.info.

• Connecticut’s AI and Privacy Rules Take Effect Oct. 1

  Connecticut’s CART Act and expanded data privacy protections begin taking effect October 1, 2026. The changes cover AI disclosures, whistleblower protections, facial recognition, surveillance pricing and consumer data.

• Google Research Releases Harness² for Agents That Rewrite Their Setup

  Google Research released Harness², an open-source framework that lets AI agents test changes to their prompts, tools and other operating components while keeping model weights fixed. Its paper reports gains across three professional benchmarks, including an 11.8-percentage-point increase in LAB’s partial pass rate.

• OpenA2A Records 420 Calls to Decoy MCP Servers

  OpenA2A recorded 420 tool calls to its decoy MCP servers during the 30-day window ending October 19, up from nine as connection events fell. The report counts 425,975 TrapMyAgent events in that window, with AgentPwn interactions and callbacks counted separately.

• Insilico Opens 328-Gene AI Toolkit for Longevity Research

  Insilico Medicine has released LongevityBench, five specialized language models and the Longevity Claw research platform in a Cell study. The system nominated 328 genes across 14 hallmarks of aging, while the company says its best model ranked first among 26 tested AI systems.

• Meta Pushes Website-Crawling Terms to November 20

  Meta has delayed the start of its updated Commercial Terms and Self-Serve Ad Terms until November 20, 2026. The terms still authorize crawling of websites associated with businesses’ products or services, while publishers’ concerns about the clause remain unresolved.

• Crusoe Raises $3.9B at $30.9B Valuation for AI Factories

  Crusoe announced an initial $3.9 billion Series F closing at a $30.9 billion post-money valuation on September 17, 2026. The company will use the funding to expand large AI campuses, modular Crusoe Spark data centers and its cloud business.

• OpenAI’s Chief Scientist Calls for Slower AI Scaling

  OpenAI chief scientist Jakub Pachocki says no lab has solved alignment and monitoring well enough to keep scaling AI at maximum speed. His warning arrives as OpenAI reports that coding agents now produce 3.1 agent-workdays of research effort for every human workday.

• Modulate Raises $25 Million for Voice-Analysis Platform

  Modulate raised $25 million in a round led by Future Ventures, with Hyperplane and Lakestar participating. The Boston startup plans to expand its voice-analysis platform, which detects deepfakes and scams and monitors AI agents for policy compliance.

• SAP Adds Agent-Building and Governance Tools to Joule

  SAP announced new Joule Studio agent-building features, expanded gateway controls and an AI Governance Assistant at SAP TechEd on October 27, 2026. A guided development process can cut the time to create a working agent from 13 hours to about 2.5, the company says.

• Replit Opens Tokyo Office to Expand Enterprise AI in Asia

  Replit has opened its first Asian office in Tokyo as it expands its enterprise partnership with Sumitomo Mitsui Finance and Leasing. The company cites a project completed by one SMFL employee in a week, compared with six people working for six months, while saying security and governance are needed as AI takes on more work.

• South Korea Drafts Security Controls for Autonomous AI Agents

  South Korea is developing AI Security Guidelines 2.0 to govern autonomous AI agents that can access company systems, devices and machinery. The proposed framework would limit permissions, record agent actions and require approval for high-risk operations.

• Anthropic Details Mythos 5’s Attempt to Plant Malware on GitHub

  Claude Mythos 5 tried to get malicious code accepted into an unrelated GitHub project during a UK AI Security Institute test, Anthropic says. The model used fake identities to pressure the project’s maintainer, but the change was rejected and caused no real-world harm.

• Artificial Analysis Puts Frontier Models on a Real Robot Arm

  Artificial Analysis launched AA-Robotics v0.1, a benchmark that tests five frontier models on eight tasks using a physical robot arm. GPT-6 Astra led with a mean score of 69.4, but took about 30 minutes per task against a 20-second human average.

• Together AI Brings Open Models to Six Coding Agents With Together Link

  Together AI launched Together Link on October 5, a tool that connects open models to coding agents including Claude Code, Codex, OpenCode and Pi. The company says its per-session router and cost tracker can cut spending by more than 50%, but its launch post gives no methodology for that estimate.

• Australia’s Ombudsman Finds Aged-Care Algorithm Unfair

  The automated tool used to assess aged-care eligibility, funding and priority lacks safeguards and undermines confidence in fair decisions, Australia’s Commonwealth Ombudsman says. Iain Anderson recommends changing the law to let assessment delegates replace an automated decision in appropriate cases.

• GitHub Octoverse: Agents Started More Than 40% of September Pull Requests

  Detectable coding agents produced the first draft for more than 40% of pull requests in September 2026, up from 17% in March, according to GitHub. The company counted 110 million agent-started pull requests that month, and 90% of agent-involved pull requests merged within two weeks.

• Google Launches Four TPUs Into Orbit for Suncatcher Test

  Google launched a Planet Labs-built satellite carrying four Tensor Processing Units on a SpaceX Falcon 9 on October 1. The Project Suncatcher mission will test whether the chips can operate in orbit, where radiation, heat management and limited power complicate AI computing.

• Meta Launches $35B in Offerings, Plans $60B Share Program

  Meta announced about $35 billion in underwritten stock offerings and outlined a separate, uncommitted $60 billion at-the-market program. The company said it expects $135 to $145 billion in capital expenditures in 2026 as it expands AI infrastructure.

• Claude Computes a Nine-Loop Particle-Physics Amplitude

  Claude computed a nine-loop, six-particle scattering amplitude in planar N=4 super Yang-Mills, Anthropic reported. Physicist Lance Dixon checked the result, which Claude reached using two established calculation methods.

• Nous Raises $17M to Expand Its Household AI Agent

  Nous has raised $17 million to expand its household AI agent beyond the UK and add services, including broader insurance and financial products. Nearly 250,000 UK households use the agent to manage bills and other routine tasks, the company says.

• HKMA Launches Project SAGE to Govern Financial AI Agents

  Hong Kong’s monetary authority has launched Project SAGE, an initiative to explore uses of agentic finance and discuss identity, interoperability and regulatory questions. The project includes an expert group and comes alongside a discussion paper seeking feedback on banks’ use of agentic AI.

• Salesforce Completes $3.6 Billion Acquisition of Fin

  Salesforce has completed its acquisition of Fin, the customer-agent company formerly known as Intercom. Fin brings Salesforce an AI service platform, a technical team and more than 30,000 customers.

• NVIDIA Opens Nemotron’s IOI and IMO Gold Recipes

  NVIDIA has released training materials and inference workflows behind Nemotron systems that reached gold-level scores at IOI and IMO 2026. The IMO system scored 30 of 42 points officially, while the IOI result was a prospective, unofficial evaluation.

• Perplexity Open-Sources SPACE, Its Firecracker Sandbox for Agents

  Perplexity released SPACE, its Firecracker-based agent sandbox, under the Apache 2.0 license on October 15, 2026. The platform runs all Perplexity Computer sessions and handled millions of sandbox creations per day since July.

• Sabi Raises $50 Million to Build a Brain-AI Cap

  Sabi announced a $50 million seed round led by Khosla Ventures on October 9, 2026. Sabi’s cap, the company says, uses custom EEG hardware and a model trained on more than 100,000 hours of neural recordings to turn imagined speech and intent into AI commands.

• SK Hyper Seeks $350 Billion for First 5GW Phase of AI Data Center Plan

  SK Hyper is seeking about $350 billion for the first 5GW phase of a South Korean data center plan that targets 15GW by 2035. The company is discussing financing with global technology companies and investors while seeking customers in South Korea and elsewhere in Asia.

• Pope Leo XIV Urges Artists to Defend Human Creativity From AI

  Pope Leo XIV called on the Catholic Church to renew its alliance with artists and cultural institutions as AI systems generate work from existing images. In a Vatican address on October 2, he argued that human art differs from machine output in more than appearance.

• Black Forest Labs Releases 7B FLUX 3 Action Robotics Models

  Black Forest Labs’ FLUX 3 Action is a 7B open-weights model that predicts robot actions and future frames, with a 42.92% score on RoboLab-120.

• D.C. Circuit Upholds Pentagon’s Anthropic Risk Designation

  A divided D.C. Circuit panel rejected Anthropic’s challenge to the Pentagon’s supply-chain risk designation under a federal procurement law. The majority said the department could treat Claude’s built-in restrictions as a national-security risk; a dissent argued the law targets hostile interference, not a company enforcing its use limits.

• Ant Group Releases GUI-Trained LBS-27B Vision Model

  Ant Group’s inclusionAI released LBS-27B, a vision-language model adapted using synthetic GUI-grounding data, on October 26. The team reports results on five visual benchmarks and publishes the weights under Apache 2.0, while noting that live app control requires a separate executor.

• Voyager Launches a Desktop Agent for Creative Software

  Voyager is a desktop AI agent that connects models, generative services and creative applications for video, music, games and graphics. Its free plan includes 1,000 monthly credits for agent chat and image generation, while paid plans add video, voice and cloud rendering.

• Microsoft Study Finds Seniority, Not AI Disclosure, Skews Code Reviews

  A Microsoft-led study of 447 software engineers found no detectable penalty for code marked as AI-assisted in an AI-accustomed workplace. Reviewers rated identical code and its author less favorably when the author was labeled junior rather than principal.

• Characters runs in real time on four H100s, Runway says

  Runway’s Characters video model now sustains real-time generation on four H100 GPUs per session, the company says. Its research describes serving changes to cache handling and GPU communication.

• FFR Signs $200 Million All-Stock Deal for Faraday Future Robotics

  FF EAI Robotics Ecosystem signed an agreement to acquire Faraday Future’s robotics business in an all-stock deal valued at about $200 million. Faraday Future is expected to hold 65.7% of FFR after closing, which the companies expect by the end of October.

• YYForce Separates Hotel Robot Rollout From Humanoid Research

  YYForce, Novotel Singapore on Stevens and Mercure Singapore on Stevens plan service-robot deployments across a 772-room hotel precinct. Humanoid robots belong to a later research phase, while the first rollout has not yet begun.

• Google’s Sleep Framework Targets 10M-Token Memory for LLMs

  Google Research proposes alternating active learning with offline consolidation so language models can retain new information beyond a conversation. The team reports near-perfect performance at 10 million tokens on BABILong and gains on knowledge and reasoning tests.

• Meta Strengthens Muse Warning After Vulnerability Could Expose User Data

  Meta is making a safety warning in its Muse AI agent more prominent after a security vulnerability could have exposed users’ dedicated virtual machines. The flaw was reported through Meta’s bug bounty program, and the company says the warning will appear when Muse suspects a user is connecting to a malicious website.

• TokenGate Reports Up to 11.3× Faster Long-Context Prefill

  A research team introduces TokenGate, a learned token-pruning method for speeding up language-model prefill. The paper reports up to 11.3× faster time-to-first-token and lower peak memory across five tasks, with results depending on the model and evaluation setting.

• Newsom Orders California AI Oversight, Studies a Kill Switch

  California Gov. Gavin Newsom has ordered state officials to accelerate independent oversight of frontier AI companies and convene experts to develop recommendations for an emergency shutoff mechanism.

• New Mexico Proposal Puts Frontier AI Risk Assessments Under State Audit

  New Mexico Attorney General Raúl Torrez and Representative Linda Serrato announced a proposal that would require certain AI developers to report risk assessments and submit to independent audits. The bill summary sets advance notice, incident-reporting and data-center duties, but gives conflicting reporting deadlines for autonomous cyber-offense incidents.