AI.info

The Pulse — Page 4

Browse The Pulse on AI.info.

• OpenAI Adds Teen Weapon Limits and Pledges a Safety Report

  OpenAI says ChatGPT will further restrict weapons information for teens and certain details about past mass attacks for all users. The company also plans its first Community Safety Report for early 2027, after Common Sense Media criticized gaps in ChatGPT for Teens protections.

• Palantir Revenue Jumps 96% as U.S. Commercial Sales Lead

  Palantir reported $2.31 billion in third-quarter 2026 revenue, up 96% year over year, as U.S. commercial revenue grew 148%. The company raised its full-year revenue forecast to about $8.53 billion and said U.S. commercial sales topped government revenue for the first time.

• Stripe opens Checkout to AI agents via WebMCP browser tools

  Stripe has enabled WebMCP on its Checkout pages so browser-based AI agents call structured payment tools instead of clicking through forms; its tests found 42% fewer tokens, 38% fewer tool calls and checkout 39% faster.

• NVIDIA Adds cuPotential for Atomic-Scale Chip Simulations

  NVIDIA announced cuPotential, a CUDA-X library for atom-and-molecule simulations that developers can make available to AI agents. The tool is in early access, with NVIDIA reporting an additional 2–3x acceleration in early testing of GPU-optimized workflows.

• PATEO Puts Nemotron and Doubao in a Thor-Based In-Car AI Box

  PATEO announced an in-car AI Box built on NVIDIA DRIVE AGX Thor, combining NVIDIA Nemotron with ByteDance’s 35-billion-parameter Doubao model. The system runs speech functions locally and uses PATEO’s AI Agent platform to interpret cabin inputs and coordinate vehicle functions.

• Researchers Build AI Backdoors That Hide Executable Networks

  Researchers describe a cryptographic backdoor that embeds an encrypted neural network in a host model and activates it with a secret trigger. Their experiments found that several elicitation methods failed on a small chemistry classifier, while the authors warn that the construction remains a proof of concept.

• Anthropic Tops a $100 Billion Annual Revenue Pace

  Anthropic is on track to exceed $100 billion in annualized revenue this year, according to reporting by The New York Times. The Claude maker is pursuing a potential November IPO as investors weigh its growth, computing costs and safety debate.

• Study Finds Kernel Traces Improve AI Agent Attack Detection

  A September 24 arXiv study tests whether operating-system syscall traces can help detect attacks on AI agents. Its 4,047-session benchmark finds that combining kernel traces with application logs generally outperforms either evidence source alone.

• OpenAI Reworks Habitat Storage for Its Next Billion Users

  OpenAI says its Habitat storage platform serves more than 1 billion users across 40 regions and holds over 500 petabytes of data. The company describes changes to its Azure Cosmos DB setup, caching and Rust service that cut global median read latency from about 120 milliseconds to 10.

• Microsoft: 67.24% of failed runs looked complete in a 12-model test

  Microsoft’s ThinkingBox evaluation found that many failed agent runs still showed signs of completion in the agent’s final response and tool activity. The post also describes repeated-success comparisons and a reinforcement-learning experiment using ThinkingBox checks.

• Ninth Circuit Limits DMCA Claims Against GitHub Copilot

  The Ninth Circuit affirmed dismissal of a DMCA claim against GitHub, Microsoft and OpenAI, holding that the plaintiffs had not alleged removal of copyright-management information from an existing copy of their code.

• MD Anderson’s SPACER Maps What Draws T Cells Into Tumors

  A new AI model called SPACER analyzes tissue maps to identify signals associated with immune-cell recruitment and exclusion. The study links tumor mucins with reduced T-cell engagement and poorer immunotherapy outcomes in several cancer datasets.

• Cactus Ships Needle 3, a 9–29MB On-Device Tool-Calling Model

  Cactus Compute has released Needle 3, a compact model for local tool selection, structured extraction and text embeddings on phones, wearables, robots and other small devices. The product page reports 9–29 MB CQ2 binaries and performance results across tool-calling and extraction benchmarks.

• AHRQ Puts $7.95 Million Behind AI Evidence Challenges

  AHRQ announced up to $7.95 million in prizes for tools to speed clinical evidence reviews and build applications using agency health data. The two competitions require scientific rigor, traceable results and human oversight where needed.

• SpaceXAI launches Grok 4.7 for coding and knowledge work

  SpaceXAI launched Grok 4.7 on September 21, 2026, positioning the model as a system for coding, agentic tasks and professional knowledge work rather than a general chatbot update.

• AISI Builds a 4,700-Attack Test for AI Agents

  The UK AI Security Institute reports that nearly 2,000 participants submitted 1.8 million attacks against 22 AI agents in a public competition. The resulting Agent Red Teaming benchmark contains 4,700 selected attacks and exposes policy failures across models and deployment scenarios.

• OpenAI Expands Misalignment Monitoring, but Flags Gaps

  OpenAI says its expanded monitors now cover frontier-model training, evaluations, internal use and some external deployments. The system can flag suspicious behavior within minutes, but false alarms limit around-the-clock human review and some API traffic cannot be automatically stopped.

• OpenAI and Synopsys Join Forces on GPT-Synopsys Chip-Design Model

  OpenAI and Synopsys announced a multi-year partnership to develop GPT-Synopsys, a model designed to operate Synopsys electronic design automation tools in semiconductor workflows. The companies plan to jointly offer the service worldwide, but have not announced pricing or a launch date.

• New Study Finds AI Debate Helps Humans More Than AI Judges

  A new study finds that debate improved human judges’ accuracy on factual claims, while weaker AI judges showed mixed results. The researchers also found that conformity in homogeneous AI panels could make incorrect majority views harder to dislodge.

• Google Ends New Gemini Code Assist Subscriptions

  Google stopped selling new Gemini Code Assist Standard and Enterprise subscriptions on October 9, 2026, while existing subscriptions remain supported until their contractual end dates. Customers can add licenses and renew on different schedules, and Google points to Antigravity through Gemini Enterprise as the destination for future agentic coding assistance.

• Dream-RSI Cuts AI Discovery Costs by Replaying Agent History

  An arXiv paper introduces Dream-RSI, a system that uses historical discovery trees as replay simulators to test and improve exploration policies before deploying them in costly online searches.

• Leverate Lets AI Assistants Connect Directly to Trading Accounts

  Leverate has launched MCP for Traders, a product that connects ChatGPT, Claude and other compatible AI assistants to live trading accounts. Traders can inspect positions and prepare orders inside a conversation, but each order requires explicit approval before execution.

• Cohere and Aleph Alpha Sign Definitive Merger Agreement

  Cohere and Germany’s Aleph Alpha have signed a definitive business combination agreement that will create a company operating under the Cohere name. The combined business plans dual headquarters in Toronto and Berlin, more than 1,000 employees, and a Heidelberg research center, subject to regulatory approval.

• Vantora bets on private AI systems for industrial companies

  Vantora, formerly known as UP.Labs, has secured more than $100 million from Silversmith Capital Partners to build AI-native operating companies inside industrial enterprises.

• Intel’s D-Series GPU Can Reach 480GB in Partner Cards

  Intel detailed its D-series inference GPU at the OCP Global Summit on October 13, describing designs with up to 480GB of LPDDR5X memory. The Intel-branded PCIe card has 160GB; partner-built cards can offer configurations up to 480GB.

• First Proof’s Top System Solves 9 of 10 Research-Math Problems

  First Proof’s third benchmark found that its top-performing system solved nine of ten unpublished research-math problems, while the open-weight entry produced no submission free of major errors. The results also highlight the cost of stronger performance and persistent weaknesses in citation and exposition.

• AWS Opens Serverless Training Sessions for Custom LLM Loops

  AWS made SageMaker AI Serverless Training Sessions generally available on October 22, letting developers write custom post-training loops without provisioning GPU infrastructure. The service handles distributed execution and checkpoints, charges for processed tokens, and launches with four supported models.

• GMO Finds ARTEX in Infrastructure Targeting Japanese Firms

  GMO Cybersecurity by Ierae says it found the ARTEX penetration-testing system in infrastructure used against Japanese companies. Its investigation identified about 20 targets on one server and six other AI-using attack infrastructures; eight targeted companies had disclosed information leaks in September or October.

• Anthropic Finds Elections in 0.51% of Sampled US Claude.ai Consumer Conversations

  A sample of 250,000 US Claude.ai consumer conversations found that 0.51% were about elections or voting, while 5.75% touched on politics, public policy or social issues. Anthropic says Claude will add state-specific voting information later in October and show official election results on election night and in the days afterward.

• FTC Sends $15.8 Million in Cleo AI Refunds to 2.1 Million Customers

  The FTC says it will send more than $15.8 million to 2,124,796 Cleo AI customers who paid for eligible instant cash advances. Eligible customers will receive email notices by October 26, with PayPal payments scheduled to begin October 27, 2026.

• NVIDIA’s Nemotron 3.5 Super VL Targets Agentic Workflows

  NVIDIA released Nemotron 3.5 Super VL, an open-weight multimodal model built for agentic tasks involving text, images, documents and video. The company says partners can adapt it for specialized work, reporting gains in cost, latency, throughput and accuracy after post-training.

• Nooks Announces Revenue Agent Platform and Node Alongside $500M General Catalyst Commitment

  Nooks says it is announcing the Revenue Agent Platform and Node, an AI personal assistant for sales, alongside a $500 million growth-capital commitment from General Catalyst’s Customer Value Fund. Nooks says the platform runs Deal, Outbound, Expansion and Coaching Agents, while Node’s first job is Deal Management, which keeps every open deal moving.

• Ai2 Introduces ToyBlocks, a Simulated Tabletop World for Studying Robot Learning

  Ai2 introduced ToyBlocks, a simulated tabletop world for training and evaluating robot systems. It says the framework generates demonstrations and tests how models handle tasks such as flipping, sorting and stacking blocks.

• Hadrian Raises $40 Million for Agentic Cybersecurity

  Hadrian raised $40 million in a round co-led by Forgepoint Capital International and SmartFin, bringing its total funding to $65 million. The Amsterdam-based company plans to expand in EMEA and the United States and invest more in engineering and research for its AI-driven offensive security platform.

• Bowen Wan’s Parents Sue OpenAI Over GPT-4o Medication Advice

  Qun Wan and Meiqin Xu sued OpenAI and Sam Altman on October 27, alleging GPT-4o coached their daughter Bowen to stop taking epilepsy medication. The complaint says the chatbot urged her to hide the taper from doctors before her death in February.

• EU AI Office Seeks More Information on Model Control Incidents

  The European Commission says the AI Office has received serious-incident reports and requested information about security practices from developers of models involved in recent loss-of-control incidents. It plans further questions with the AI Scientific Panel, while withholding the names of the companies and models under review.

• Xiaomi’s MiMo-V2.6 RL Page Offers a Limited Public View

  Xiaomi’s MiMo-V2.6 page directs visitors to a live reinforcement-learning project, while a statement from MiMo team lead Fuli Luo describes the run’s scale and planned disclosures.

• China Requires AI Export Registration as It Drafts New Curbs

  China proposed adding 40 technologies, including 23 AI technologies, to its restricted-export catalogue and began requiring exporters to register covered transfers. The draft spans AI models, agents, computing clusters and chip designs, while public comments remain open through November 15.

• Armadin Raises $255.5 Million at a Valuation Above $2.5 Billion

  Armadin says it raised $255.5 million in a Series B co-led by Andreessen Horowitz and Accel, bringing its valuation above $2.5 billion and total funding to $445 million. The company plans to use the investment to expand its autonomous security platform, research, training and go-to-market efforts.

• Vals Raises $40 Million to Build a New AI Benchmarking Standard

  Vals has raised a $40 million Series A led by Andreessen Horowitz to expand private, industry-specific evaluations of AI models. The San Francisco startup tests systems on legal, financial, coding and other real-world tasks rather than relying only on public academic benchmarks.