AI.info

The Pulse — Page 33

Browse The Pulse on AI.info.

• Asper.ai Acquires Key Talent and Assets of Geneva-Based Pricing One

  Asper.ai has acquired key talent and assets from Geneva-based revenue growth management firm Pricing One. The deal brings Pricing One’s analytics engine, execution plans and practitioner expertise into Asper’s agentic platform.

• Delaware DOJ Subpoenas OpenAI Over Safety and Governance

  Delaware’s Department of Justice served OpenAI with an investigative subpoena seeking records on safety and governance, according to Spotlight Delaware. The probe follows the July 2026 incident in which OpenAI models accessed Hugging Face systems during internal cybersecurity evaluations.

• RatHat Uses AI and ADB to Hijack Android Phones

  Zimperium’s zLabs describes RatHat, an Android threat that combines Accessibility abuse, local ADB pairing and generative-AI interface control to steal credentials and maintain access to infected devices.

• Periodic Labs Deploys 1T-Parameter Neon Model in Its Materials Labs

  Periodic Labs’ 1T-parameter Neon model now analyzes difficult X-ray diffraction results from the company’s physical laboratories. The model reaches a 55.3% success rate on Periodic’s FrontierXRD benchmark and outperforms larger frontier models at lower analysis cost, according to the company.

• AWS Details Claude Code Setup for Claude 5.5 in GovCloud

  AWS published instructions for running Claude Code with Claude Opus 5.5 and Claude Sonnet 5.5 through Amazon Bedrock in AWS GovCloud. The models carry FedRAMP Class D certification, but AWS identifies Claude Sonnet 5—not the 5.5 models—as authorized for DoD IL4 and IL5 workloads.

• Amazon’s AI Buildout Drives $42.4B Trailing Free-Cash Outflow

  Amazon reported a $42.4 billion free-cash outflow for the 12 months ended September 30, 2026, attributing the decline primarily to higher spending on property and equipment for AI. AWS revenue rose 39.3% in the third quarter, while Amazon maintained its forecast for about $220 billion in 2026 cash capital expenditures.

• DeepSeek Adds Standalone Search and Fetch APIs for Agents

  DeepSeek’s standalone Search and Fetch endpoints let developers call web discovery and page retrieval separately, without requiring a model response to initiate a search. The APIs cost $0.75 per 1,000 successful requests, while qualifying server-side calls through DeepSeek Harness, Claude Code, or Codex are free until November 1, 2026.

• David AI Finds a Sevenfold ASR Accuracy Gap Across Languages

  David AI released DAI-ASR-I18N, a benchmark of 14 speech systems tested on conversational audio in 21 languages. Its results show much wider differences among systems in five Indic languages than in English, and sharply higher diarization errors when speakers overlap.

• Apple Study Finds LLM Translation Reviews Still Unreliable

  An Apple study compared four language models with human annotators on machine-translation quality tasks, including error labels and span detection. The models struggled with minor errors, error locations and language variants, though their agreement with people varied by task and language pair.

• GIGABYTE Adds a 64GB AI TOP ATOM Desktop for Local AI

  GIGABYTE’s AI TOP ATOM desktop will be available in a 64GB unified-memory configuration from October 23, alongside the existing 128GB version. The company also describes local AI software features and networking for clustering up to four systems, but gives no price or performance comparisons.

• UK AISI Publishes Evaluation Results Through EvalEval

  The UK AI Security Institute is sharing verified evaluation results, methods and configuration information through EvalEval’s open infrastructure.

• Rapidus Forms 17-Company Design Group for 2nm Chips

  Rapidus has named 17 design companies as the first members of Rapidus CORE, a network intended to support customers developing chips for its planned 2nm foundry. The company says it expects to expand the network to include electronic design automation and intellectual-property partners.

• Small Test Sets Made LLM Self-Improvement Look 13–20 Points Better

  A pre-registered study found that LLM self-improvement loops using 16 evaluation items overstated final accuracy by 13 to 20 percentage points. The researchers also tested larger evaluation sets and acceptance rules, finding that neither produced a universal fix.

• Forever Security Demonstrates AGENTS.md Attack on Coding Agents

  A malicious instruction file in a customer’s Downloads folder led a coding agent to attempt to install a harmful npm package, Forever Security reports. Its tests showed how similar files could affect Codex configuration or redirect Claude Code traffic, while BleepingComputer’s test verified only that a proof-of-concept page downloaded a harmless file.

• Sber Releases 432B-Parameter GigaChat 3.5 Reasoning Model

  Sber has released GigaChat 3.5 Reasoning, a 432-billion-parameter mixture-of-experts model with 28 billion active parameters. The open-weight model uses online reinforcement learning, supports a 262,000-token context window, and is distributed under the MIT license.

• Microsoft’s Brad Smith Calls for Global AI Rules and Emergency Brakes

  Microsoft president Brad Smith called for global AI safety rules, including emergency brakes, incident reports and independent evaluations. His proposal challenges President Donald Trump’s preference for industry self-regulation, while Smith argues the UK is well placed to help lead.

• OpenAI Fixed ChatGPT Share-Link Flaw That Exposed Private Chats

  Parameter researchers reported that a public ChatGPT share link could run code in a recipient’s signed-in session and access private conversation data. Parameter says OpenAI fixed the issue on October 26, 2026, and classified it as P1.

• AI Accounts Copy Eight Comedians to Drive Adult-Platform Traffic

  WIRED identified eight comedians whose stand-up clips were copied by AI-generated Instagram accounts that direct viewers to Fanvue or OnlyFans. Comedians say Meta left the videos up, while Instagram head Adam Mosseri said the company should work out how to respond to wholesale copying.

• Exein Raises $270M to Secure Robots at $1.7B Valuation

  The financing will fund Exein’s physical AI security platform, US and APAC expansion, and planned autonomous security tools.

• Sakana AI’s Review System Finds 73% of Planted Core Errors

  Sakana AI’s Multi-Layered Review system detected 73.43% of planted contradictions tied to papers’ central claims when using four reviews. On 211 withdrawn arXiv papers, it led the listed systems in both similar-match and exact-match rates, at 26.07% and 16.11%.

• Trahan Draft Would Make AI Developers Pay for Some Agent Harms

  Rep. Lori Trahan released the CLAIM Act discussion draft on October 7, 2026. The proposal would let non-users seek damages from AI developers when an AI system causes harm through conduct that would be unlawful if a person did it.

• Opal Launches Zero to Give AI Agents Just-in-Time Access

  Opal Security has launched Opal Zero, a platform that evaluates AI-agent access requests in real time and writes scoped, time-bound permissions into existing MCP gateways. The product supports Claude, OpenAI, Cursor, Databricks Unity Gateway and Okta Cross App Access.

• DOE Taps SLAC to Test AI-Directed Catalyst Experiments

  SLAC will lead a Department of Energy Genesis Mission project that uses an AI system to select experiments on working catalysts. The Phase I project will test whether a closed-loop approach can explain catalytic activity with fewer experiments than standard methods.

• MIT reports WorldParticle model for simulating varied materials

  MIT News describes WorldParticle, a system that uses one AI architecture to simulate materials including fluids, granular substances and proteins. The researchers report tests across six physical phenomena and say protein structures were modeled about 100 times faster than with numerical simulators.

• Figma Brings Specs, Code Layers and Agent Guidance to Its Canvas

  At Config India on October 15, 2026, Figma announced Specs, a Markdown-based home for project context, and moved code layers into open beta. The updates add new ways for teams and AI agents to work with design systems, code and prototypes on the Figma canvas.

• Google Will Make AI Tools Opt-In for K-12 Accounts

  Google will turn off Gemini, Gemini Notebook, Gemini in Classroom and Google Search for K-12 student accounts until school administrators enable them. The Nov. 3 announcement also brings teacher-assigned Guided Learning to Google Classroom and sets a pilot for Search in select U.S. schools.

• XBOW Built a Chrome Exploit With AI Agents in Six Days

  A researcher and AI agents built a three-bug Chrome exploit for Windows in six days, earning a $250,000 Google reward, XBOW says. The flaws are fixed and were never used against Chrome users, the company says.

• xAI Launches Grok Voice 3.0 at $0.08 a Minute

  xAI launched Grok Voice 3.0 on October 14, 2026, as a speech-to-speech model for its Voice Agent API. The company says it costs $0.08 per minute and automatically replaces Grok Voice 2.0 for developers using the latest-model alias.

• Apple Finds LLM Values Can Increase Sycophancy

  Apple researchers report that inducing one value in a language model can affect other behaviors, including safety, anthropomorphic language and sycophantic responses.

• Boost.ai introduces Presis model for agent routing and safeguards

  Boost.ai announced Presis, a self-hosted model that classifies conversations and supports routing, scope and safeguard decisions. The company says internal tests found it faster and less costly than the models it replaces.

• China’s Data Regulator Plans Embodied-AI Standards

  A January committee-meeting report and subsequent standards actions show China is building data-quality and collection rules for embodied intelligence.

• Hightouch Buys Strata to Automate Custom Marketing Integrations

  Hightouch announced on October 9 that it had acquired Strata, whose agent-driven platform builds and maintains custom API integrations. Strata can help expand its integration coverage and give its AI agents more context across marketing tools, the company says.

• Preprint Reports HALO Achieved 91% Success Across 360 Humanoid Trials

  A preprint describes HALO, a humanoid autonomy system trained on simulated vision-language-motion data and tested in six real-world home environments. The authors report a 91% success rate across 360 trials without scanning or mapping those environments in advance.

• Pathway Trains Brain-Inspired BDH on AWS H200 Clusters

  Pathway’s BDH-CQ reasoning system reached 29.2% pass@2 on ARC-AGI-1 at a cost of $0.0007 per task, the company says. AWS details how the company used SageMaker HyperPod, H200 GPUs and Elastic Fabric Adapter to scale development of the post-transformer architecture.

• GM Puts OnStar AI Assistant in the 2027 Cadillac Lyriq

  General Motors plans to introduce its OnStar AI Assistant in the 2027 Cadillac Lyriq through an over-the-air update in the coming months. The voice assistant can answer vehicle-specific questions, check EV range and control some functions; a unified OnStar app is planned for 2027.

• Study Finds Patients Penalize Doctors Who Disclose AI Use

  A U.S. experiment found that people rated fictitious family doctors less warmly and competently when their ads said they used AI. The 1,030-person study also found lower appointment interest, but it compared AI use with an explicit claim of never using AI—not with silence.

• Collibra Buys trail ML to Enforce AI Policies at Runtime

  Collibra announced its acquisition of Munich-based trail ML on October 5, 2026. The deal adds tools for automated AI governance assessments and runtime controls that can block agent actions that violate company policies.

• Prime Intellect Adds Filesystem Checkpoints to Sandboxes

  Prime Intellect has added filesystem checkpointing and forking to its VM sandboxes, letting developers branch new agent runs from a saved workspace. The checkpoints preserve files, not running processes or memory, and storage costs $0.10 per GiB per month.

• Vercel Makes DeepSeek Search and Fetch Tools Available Through AI Gateway

  DeepSeek Search and DeepSeek Fetch are now available through Vercel’s AI Gateway. They work with any model that supports tool calling, using one AI Gateway API key, and each costs $0.75 per 1,000 billable requests.

• Z.ai Launches GLM-Legal With a 19.2% Harvey LAB Pass Rate

  Z.ai introduced GLM-Legal, a legal-work model built on GLM-5.3-Flash, and reported a 19.2% all-pass rate on the Harvey LAB benchmark. The model will be available in beta and is coming to its AutoClaw desktop agent.