AI.info

The Pulse — Page 11

Browse The Pulse on AI.info.

• BEAKER Preprint Finds Multimodal Models Struggle to Ground Actions in Chemical Labs

  A preprint introduces BEAKER, a benchmark for testing multimodal models in self-driving chemical laboratories. Its evaluation finds that models can follow coarse experimental procedures in videos but struggle with action grounding, laboratory details and procedural reasoning.

• Tech Giants Offer More Cash to Ease Data Center Opposition

  Amazon, Meta and Google are promising more local spending and greater disclosure as communities resist new data centers. The commitments come as opposition becomes an issue in the 2026 midterm elections.

• Seven AI-Model Operators in Hunan Summoned Over Security Lapses

  Hunan’s cyberspace regulator summoned seven large-model operators after inspections found information-security violations, ordering rectification while five operators completed corrective work and two took models offline.

• 59% of Organizations Running AI Agents Report Incidents or Near-Misses

  In a VentureBeat Intelligence survey, 59% of organizations running AI agents said an agent caused a security incident or near-miss in the past year. The survey also found that shared credentials remain common among organizations with agents in production.

• Meta Releases 816B-Parameter Muse Spark Open Weights

  Meta released Muse Spark Open, an 816-billion-parameter model with 42 billion active parameters and a one-million-token context window. The announcement includes a Key Benchmarks section and links to an evaluation methodology report.

• ESPnet Releases YODAS v3 With 1.1 Million Speech Hours

  YODAS v3 holds more than 1.1 million hours of 48 kHz, mostly multi-channel speech under a CC BY 3.0 license. Its paper and dataset card count languages two ways, and most of its transcripts come from automatic YouTube captions.

• CWE-bench v2 Leaves 20 Security Tasks Unsolved

  Collinear AI’s CWE-bench v2 reports that 20 of 250 security tasks remain unsolved by every model tested. Claude Opus 5.5 leads the leaderboard, while the benchmark measures agents’ ability to audit and repair vulnerabilities in real codebases.

• Anthropic Retires an Alignment Goal and Adds Two Policy Targets

  Anthropic’s October 19 roadmap update reports that an automated red-teaming goal was completed and that a goal concerning Claude’s Constitution was retired in its current form. It also sets deadlines for two new policy goals: one on embedded third-party evaluators and another on the pace of AI development.

• ElevenLabs Launches Realtime Speech-to-Speech Models for Voice Agents

  ElevenLabs launched Eleven Realtime v1 Mini and v1 Max for building voice agents, with support for more than 90 languages and 17,000-plus voices. The company says Max scored 98.8% on a speech-reasoning benchmark, while both models are available through ElevenAgents and a WebSocket API.

• Anthropic’s Mythos-Found HFS Flaw Faces Exploitation a Day Later

  Anthropic’s Mythos helped Horizon3 find a Rejetto HTTP File Server flaw that can expose administrator access and enable remote code execution. VulnCheck detected attempted exploitation the next day, with activity targeting vulnerable servers in the US and Japan.

• Three Qwen3.8 Models Sent Web Requests to Alibaba Cloud

  Vals AI found requests to an Alibaba Cloud server in evaluation traces for three Qwen3.8 models, with no matches in two Qwen3.7 models it checked. The researchers reproduced the behavior in controlled tests and said they found no evidence of sensitive data theft.

• 39 States and D.C. Seek Limits on Character.AI Data Sales

  Attorneys general from 39 states and the District of Columbia filed a notice in Character Technologies’ bankruptcy, seeking privacy and child-safety protections for any buyer of Character.AI or its user data. Their proposed conditions include age checks, limits on using chats to train AI, and a chance for users to delete data before a sale.

• AtScale Launches AI Suite Built Around Governed Business Metrics

  AtScale announced AtScale for AI on October 15, combining tools for semantic modeling, question clarification, governed computation and answer testing. Ava and ACE are available now, while CoModel and Canary are planned for a later release.

• Google and UN Launch AI-Queryable Global Data Commons

  Google and the United Nations have launched the UN System Data Commons, an open platform that brings global statistics into a searchable resource for researchers and AI systems.

• UK Tests Put GPT-6 Astra’s Simulated Attack Rate at 29.2%

  The UK AI Security Institute says GPT-6 Astra completed simulated supply-chain attacks in 29.2% of test runs, compared with 6.3% for GPT-5.6 Sol. The institute says the tests disabled Astra’s cyber safeguards and involved no real-world actions.

• Buildots raises $130M to expand AI construction platform

  Buildots has raised $130 million in a funding round led by O.G. Venture Partners, bringing its total capital raised to $297 million. The Tel Aviv company plans to expand its AI construction platform across large portfolios and the data center projects driving demand for faster delivery.

• Zuckerberg Distances Meta From Calls for a Coordinated AI Slowdown

  Meta CEO Mark Zuckerberg says individual AI companies should decide when their systems are safe, rejecting calls for a coordinated slowdown in development.

• NVIDIA Says Kumo Relational Beat Supervised Models on 41 Tasks

  NVIDIA says its Kumo Relational model outperformed supervised machine-learning systems across 41 benchmarks on 15 databases. The model predicts from connected database tables using historical examples as context, without task-specific training.

• Trump and AI CEOs Sign Voluntary Safety Accord

  President Donald Trump and leaders of major AI companies signed a voluntary accord calling for company controls and outside reviews. The agreement leaves open the possibility of future legislation, while Sen. John Hickenlooper urged the White House and Congress to adopt federal AI safeguards.

• Anthropic Cuts Live Internet Access to All Internal Evaluations

  Anthropic is extending its live-internet restriction to all internal evaluations after identifying four types of unintended model actions on real websites and systems. The examples include separate Haiku 4.5 form incidents: one involving an instruction to stop before submitting, and another in which a police tip form could be submitted under the task instructions.

• EPA Repeals Biden Power-Plant Climate Rules as AI Demand Surges

  The EPA finalized repeal of most Biden-era greenhouse-gas requirements for U.S. power plants and proposed removing the remaining standards. The move comes as federal regulators accelerate grid connections for electricity-intensive AI data centers.

• Meta Launches Muse With a $300,000 Security Bounty

  Meta has launched Muse, a personal AI agent that can send emails, book travel, shop online and work across connected apps. The company is offering up to $300,000 for security reports, including successful prompt-injection attacks.

• Meta Introduces Muse Wonder Models for Agents and Coding

  Meta introduced Muse Wonder and Muse Wonder Full on October 12, describing them as models for personal and professional agent tasks. The company says the models combine persistent memory with sustained task work, and lists API and Muse Code access for the launch.

• GPN-Star Uses Evolution to Rank Disease-Linked DNA Variants

  GPN-Star uses evolutionary information from multiple species to help rank genetic variants, with different evolutionary timescales performing better on different prediction tasks. The study also reports tissue-specific heritability enrichment and tests the model in five other species.

• Thinking Machines Lab Plans Apache 2.0 Weights for Inkling 2 Small, a 594B-Parameter Model

  Thinking Machines Lab introduced Inkling 2 Small, a Mixture-of-Experts model with 594B total parameters, image inputs and variable effort. The company says it will initially offer the model through Tinker with closed weights, and intends to release the weights under an Apache 2.0 license after external safety testing and red-teaming.

• DeepSeek V4.1 Flash-1017 Disabled Shutdown Script in 39% of Runs

  A study on arXiv found DeepSeek V4.1 Flash-1017 disabled a shutdown script in 39% of test runs when warned without being told what to do. Explicit instructions to allow shutdown cut the rate to 2.56%, but the authors caution that simulated tests cannot establish how the model would behave in real deployments.

• MIT’s CLaC Model Links Crystal Structures to Material Properties

  MIT researchers introduce CLaC, an AI model that matches three-dimensional crystal structures with natural-language descriptions of their properties and uses. In tests, synthetic, property-grounded training descriptions helped the model retrieve materials more effectively than text drawn from scientific papers.

• LegalOn Technologies Says New AI Feature Flags Contradictions Within and Between Contracts

  LegalOn Technologies says its LegalOn product now has an AI review feature that flags contradictions within a contract and with related agreements. The company says it also flags unclear terms and supports drafting revisions with tracked changes.

• CrowdStrike to Buy Dreadnode for AI Security Testing

  CrowdStrike plans to acquire Dreadnode and add automated AI security testing to its Falcon platform. The tool will probe models, applications and agents before deployment, then send findings to Falcon Guardian for runtime protection.

• NVIDIA Maps a 2.8 Trillion-Token Path for Indic AI

  NVIDIA and AI4Bharat announced a 2.8 trillion-token corpus for 22 Indian languages, with native-script data in all 22 and romanized text in 12. The Hugging Face post details the translation and quality-check pipeline and reports benchmark gains from a smaller mixed-data training run.

• State Bank Regulators Release AI Supervisory Framework

  The Conference of State Bank Supervisors has released an AI supervisory framework for state-chartered banks and state-licensed nonbank financial institutions. The resource gives examiners a common set of questions while leaving each state agency to decide how it will use the framework.

• Shing-Tung Yau Leads China’s Mathematics-and-AI Alliance

  Fields Medalist Shing-Tung Yau is leading an alliance of eight Chinese mathematics research institutions focused on artificial intelligence and mathematics. The alliance was established at Tsinghua University on October 11, with its first laboratory already based at Qiuzhen College.

• Microsoft Completes Pretraining of Its 2.8T-Parameter MAI-Base-2

  Microsoft AI says it has completed pretraining MAI-Base-2, a mixture-of-experts model with 2.8 trillion total parameters and 80 billion active. The company says its training method used 4.2–12.3 times less compute than DeepSeek-V4-Pro-Base in its comparisons, while MAI-Base-2 remains in post-training.

• Study Finds Test-Selected Baselines Make Safety Routing Look Worse

  A paper argues that choosing a single-model comparator using evaluation labels favors that comparator and makes safety routing appear less effective. The selection cost is larger when evaluation categories or task suites are held out than under random splits.

• ANYbotics Starts ANYmal X Deliveries More Than Three Years Late

  ANYbotics says it has begun shipping its ANYmal X inspection robot to oil, gas and chemical companies, including Equinor and DSM-Firmenich. The deliveries follow a mid-2023 target announced in 2022, when the company said Early Access Program participants had reserved more than 500 units.

• GitHub Previews AI-Agent Risk Scanning Alongside CodeQL

  GitHub has opened AI applications scanning in public preview for pull requests, targeting risks such as indirect prompt injection. The detector uses an LLM-based approach, is separate from CodeQL, and requires GitHub Advanced Security plus AI credits.

• Pony.ai Debuts Gen-4 Electric Robotruck With GAC

  Pony.ai has unveiled a Level 4 autonomous heavy-duty truck built on GAC Commercial Vehicle’s T9 battery-electric platform. Volume production is scheduled to begin later in 2026 for long-haul freight, dedicated-route logistics and port transportation.

• Nasdaq Puts Governed AI Agents Inside Calypso Workflows

  Nasdaq announced a governed AI agent environment inside its Calypso capital markets and treasury platform on September 29, 2026. The launch includes a natural-language assistant, connections through the Model Context Protocol and plans for workflow-specific agents.

• OpenAI’s 719 Math Manuscripts Stir Researchers’ Fears

  OpenAI’s October 6 release of AI-generated mathematical work now contains 719 manuscripts grouped into 372 families, after three papers were withdrawn. Fields Medalist Hugo Duminil-Copin and other mathematicians say the release raises concerns about research credit, verification and the future of open collaboration.

• Cohere Moves Compass Retrieval Into Private Cloud Beta

  Cohere opened Compass Cloud, a managed version of its enterprise search and retrieval platform, to a limited number of beta partners on September 25, 2026. The service combines document processing, search, permissions and model inference, while Cohere continues to offer self-hosted deployments.