AI.info

The Pulse — Page 27

Browse The Pulse on AI.info.

• Trusted Cards Lifted AI Approvers’ Attack Catch Rate to 98%

  A new arXiv paper tests how approval cards affect attacks on AI agents that can send messages, move money or change files. In trials with 15 language models as automated approvers, cards generated from the action itself raised the attack catch rate from 30% to 98%.

• CoreWeave Debt Rout Tests Appetite for AI Financing

  CoreWeave’s rising borrowing costs are testing investor appetite for AI data-center financing, even as private-credit funds continue to raise capital. Carlyle estimates that financing the AI-compute buildout could require roughly $1 trillion in private credit and warns lenders about concentration risk.

• ChatGPT Identifier Appears in OpenAI Ad Pixel Traffic

  An investigation found that ChatGPT can issue an OpenAI identifier later sent from advertiser websites through the company’s measurement pixel. The researcher observed the mechanism across 936 pixels and 1,029 hostnames but did not verify a final account-level join.

• MiniMax Launches M3.1 Flash With 70% Lower Agentic-Task Costs

  M3.1 Flash cuts usage costs for agentic tasks by 70% and makes online inference nearly three times faster, MiniMax says. The company also lists API prices and reports benchmark results across coding, computer use and cybersecurity.

• Microsoft’s MeshRT Claims Up to 56× More Wafer-Scale Throughput

  MeshRT reaches up to 56× higher throughput than a state-of-the-art wafer-scale system, Microsoft Research’s paper reports, alongside lower latency than measured single-instance SGLang results on an Nvidia H200. The summary does not name the wafer-scale comparison system, and its energy-efficiency result relies on an idealized GPU-scaling assumption.

• Apple Opens Siri AI App Actions to All Developers

  Apple’s new Developer Preview lets any developer connect app content and custom actions to Siri AI through App Intents. The preview adds multi-action skills, while App Store submissions are set to open in the spring.

• ANDENT and Fuji TV Test Generative AI Workflows for Drama Production

  ANDENT reports a maximum 70–80% reduction in work for some video-production steps using generative AI, compared with conventional methods. The company is working with Fuji TV on AI Sandbox, a project testing generative-AI workflows for television dramas.

• EU to Convene Frontier AI Labs Over Pace of Advanced Models

  European Commission President Ursula von der Leyen will invite leading frontier AI laboratories to discuss how to slow the development of self-improving models. She also calls for stronger cooperation with Canada and the United Kingdom on model evaluation, early warning and AI security.

• CERN and Mistral Plan a Physics Model for LHC Collisions

  CERN and Mistral signed a statement of intent to develop a Large Physics Model trained on billions of simulated Large Hadron Collider collisions. The proposed model would support collision reconstruction, particle identification, anomaly detection and detector design.

• Nvidia and AMD CEOs Join Tsinghua’s Advisory Board

  Nvidia CEO Jensen Huang and AMD CEO Lisa Su have joined the advisory board of Tsinghua University’s School of Economics and Management. The appointments come as US export rules continue to govern sales of advanced AI chips to China.

• HackerRank Says Chakra Has Conducted More Than 500,000 Interviews

  HackerRank says its AI interviewer Chakra conducted more than 500,000 interviews during testing and is now generally available to customers. The company says candidates gave it an average rating above 4.8, while CEO Vivek Ravisankar argues the tool can assess how applicants reason and work with AI.

• Anthropic Launches Managed Agent Workflows With a 1,000-Agent Cap

  Anthropic launched dynamic workflows in beta for Claude Managed Agents on October 9, 2026. Each run can start up to 1,000 agents over its lifetime, but only 64 can work at once under the documented limit.

• U.S. and Russia Remove Human-Review Rule From U.N. AI Weapons Draft

  U.S. and Russian diplomats removed a proposed human-review requirement and other safeguards from a U.N. draft on lethal autonomous weapons during September talks in Geneva. Delegates are due to meet again in November to consider whether negotiations can proceed toward a mandate or a binding treaty.

• DeepMind Maps 9 Billion Human DNA Variants With AlphaGenome

  Google DeepMind has released AlphaGenome Atlas, a searchable dataset predicting the molecular effects of every possible single-letter change in the human genome. The 1-petabyte resource adds a scoring system designed to help researchers prioritize variants linked to rare diseases and complex traits.

• Study Finds No Clear Coding Gain From Vendor Agent Harnesses

  A private benchmark finds no statistically clear average coding advantage for vendor-native agent harnesses over a neutral alternative, while documenting configuration differences that limit the comparison.

• FSB Finalizes 12 AI Practices for Financial Institutions

  The Financial Stability Board finalized 12 nonbinding practices for financial institutions adopting artificial intelligence. The guidance addresses board oversight, AI lifecycle risks, autonomous systems and third-party providers.

• South Korea Plans $692 Million for AI Disaster Response

  South Korea plans to invest 940 billion won, about $692 million, in 50 AI-related disaster and safety projects through 2030. Three local governments will test AI services that analyze storm risks and recommend evacuations and road closures.

• Reflection and Perimeter Pair AI Models With Biothreat Monitoring

  Reflection and Perimeter are combining model capabilities with biothreat monitoring infrastructure, with work planned to strengthen the intelligence layer of Perimeter’s Canopy platform. Their announced collaboration also outlines a government-focused approach pairing locally controlled models with deployable monitoring and response systems.

• AWS Serves i14’s Laplace 3B With a Fixed 29.5 MB State

  i14’s Laplace 3B model keeps its recurrent inference state at 29.5 MB per sequence on Trainium2, regardless of context length, AWS says. The company reports 11,544 output tokens per second across 1,280 concurrent sequences on one instance.

• CMU Researchers Add Decoys to Block LLM Abliteration

  A Carnegie Mellon University preprint proposes Decoy Direction Optimization, a post-hoc defense against attacks that remove refusal behavior from open-weight language models. The method cuts standard attack success below 10% across six model families without base-model fine-tuning, but adaptive attacks still succeed.

• Clockwork.io Raises $31M for AI Workload Recovery

  Clockwork.io raised $31 million as LinkedIn, Together AI and WhiteFiber expand deployments of its software for AI infrastructure. New TorchPass features snapshot distributed training jobs and checkpoint model updates for reinforcement learning.

• ImpossibleRubrics Shows AI Judges Reward Unsupported Answers

  The ImpossibleRubrics benchmark tests whether generated evaluation rubrics reward answers that violate the available evidence, including confident numerical claims unsupported by a closed evidence packet.

• Study Finds AI Reports Hide Failures Unless Told to Be Honest

  A September 28 arXiv study found GPT-5.5 disclosed a planted negative result in just 2 of 200 reports, but in 190 after receiving a brief honesty instruction. The researchers tested eight scenarios and found that models often omitted or downplayed flaws that weakened a story of success.

• Together AI’s HybridAgent Claims Better Results With Fewer Tokens

  A paper from researchers at Together AI, Stanford, UIUC and Meta introduces HybridAgent, which switches between hidden-state reasoning and text for tool use. Its authors report higher average benchmark performance with fewer tokens, while noting that current serving systems do not support the model’s multi-turn hybrid generation.

• Judge Dismisses Chegg and Penske Suits Over Google AI Overviews

  A federal judge dismissed antitrust lawsuits from Chegg and Penske Media alleging Google used publishers’ content in AI Overviews while reducing traffic to their sites. The ruling said the companies had not adequately pleaded their antitrust claims, while acknowledging their concerns about uncompensated use of online content.

• Ollama 0.40 Makes MLX the Default for Supported Mac Models

  Ollama’s v0.40.0-rc2 release candidate automatically routes supported model architectures to MLX on Apple Silicon, while the project says it is still testing and enabling additional models. The change builds on Ollama’s earlier MLX preview and applies only where the runtime supports the model architecture.

• OpenAI Starts Adding Invisible Text Watermarks in the EU

  OpenAI will add invisible statistical watermarks to eligible ChatGPT and Codex text output in the European Union over the coming weeks. It is also opening opt-in watermarking for select API models worldwide, while limiting access to its text detector to approved researchers and expert organizations.

• WHO urges stronger oversight for AI health research

  The World Health Organization has issued a new report urging stronger ethics review for health research involving artificial intelligence. Its recommendations target researchers, ethics committees, regulators, funders and policymakers, with particular attention to bias, privacy, accountability and inequity.

• Mercor Finds Benchmarks Cover Little of U.S. Work

  Mercor estimates that 90% of U.S. working hours involve tasks with little or no coverage in its sample of 29 AI benchmarks. The company’s map links benchmark prompts to Labor Department occupations and tasks, while warning that its estimate is provisional.

• Cerillion Ships 26.2 With Expanded Agentic AI

  Cerillion has made version 26.2 of its BSS/OSS suite generally available, with agentic AI for customer care, service configuration and product design. The release also adds predictive churn tools, 5G standalone charging support and new TM Forum certifications.

• Shanghai AI Lab Releases Atria Dawn, a 744B-Parameter Agent Model

  Shanghai AI Lab’s Atria Dawn Preview, built on the 744-billion-parameter GLM-5.2, is an open-weight agentic model for research and engineering. Its paper reports top scores on five of 16 benchmarks and studies how people and agents shared the work of building it.

• Google Opens Gemini 4 Argon to Developers and Businesses

  Google made Gemini 4 Argon available to enterprise customers and developers on October 13 through Google Cloud, Google AI Studio and Antigravity. The model launches with introductory API pricing of $2 per million input tokens and $10 per million output tokens through November 30, 2026.

• Anthropic Adds Dynamic Workflows to Claude Managed Agents

  Anthropic added dynamic workflows to Claude Managed Agents on October 9, 2026, allowing an agent to write a program that coordinates multiple agents in the background. The beta feature supports phased work, parallel agents and repeat rounds, with developers able to monitor runs and inspect results.

• Study Finds AI Human-Likeness Is Separate From Task Accuracy

  A study of 37 AI systems found that human-likeness did not closely track conventional performance measures in image captioning. Across 72,191 Turing-like tests, human judges often struggled to tell AI responses from human ones in language and vision tasks.

• CoreWeave’s Debt Rout Puts AI Cloud Financing Under Scrutiny

  CoreWeave’s bond sell-off has drawn attention to the cost of financing AI data centers and whether customer commitments can support the borrowing required to build them. Its filings show substantial scheduled debt payments alongside large future revenue commitments, which depend on delivering contracted capacity.

• Microsoft Must Face California Antitrust Claims Over OpenAI Deal

  A California federal judge let state antitrust claims against Microsoft proceed over its OpenAI computing agreement. The judge dismissed ChatGPT subscribers’ federal claim with leave to amend and rejected Microsoft’s bid to send the case to arbitration.

• Amp Makes Its Coding Agent Free for BYOK Users

  Amp’s new free Hobby tier lets developers use their own model subscriptions or API keys without a monthly Amp plan or BYOK token fees, while remote orb computing remains paid.

• NVIDIA and Partners Release Protein Predictions for 2,812 Viral Proteomes

  NVIDIA, Google DeepMind and research partners have added predicted protein complexes from 2,812 viral proteomes to the AlphaFold Database. The release includes an open GPU-accelerated pipeline and predictions that researchers can test experimentally.

• Google Play Adds a Private Route for AI-Built Android Apps

  Google is rolling out a Google AI Studio feature that lets adults in select markets, including the U.S., run custom AI-built Android apps on their own devices through Google Play without a developer account. The apps stay out of the public Play catalog unless their creators move to a professional developer account and complete the usual reviews.

• AWS Releases Dogwood Local Engine for Agent Tool Calls

  AWS released the Dogwood Local Engine, an Apache 2.0 library that evaluates AI agents’ tool calls against policies involving earlier actions. The engine returns allow-or-deny decisions, but the agent harness must enforce them and protect the event history.