AI.info

The Pulse — Page 60

Browse The Pulse on AI.info.

• xAI to Watermark Seven Grok API Models by Default on Nov. 7

  xAI says invisible watermarks will be added by default to text from seven Grok API models starting November 7, 2026. Team administrators can opt out per model, while the API adds an x-watermark response header and requires no request-code changes.

• OpenAI Trains GPT-6 Astra on Ironclad Contract Workflows

  OpenAI and Ironclad built 11 contracting tasks to train and evaluate OpenAI’s GPT-6 Astra agent. Astra scored 55% on the research evaluation, compared with 41.6% for GPT-5.6 Sol, but OpenAI says the results do not measure customer time savings.

• Mohiuddin Finds MCP Security Bugs Across Five Organizations

  Independent researcher Syed Anas Mohiuddin says security teams at Google, JPMorgan Chase, Weaviate, France’s DINUM and Indonesia’s Tangerang City fixed MCP-related server-side request forgery bugs. He distinguishes those confirmed fixes from five U.S. federal findings that were still under review.

• Vercel adds MiniMax H3 Image models to AI Gateway

  Vercel has made MiniMax H3 Image and its Highspeed variant available through AI Gateway. Both can generate and edit images, with different output-token prices.

• OpenAI Withdraws Three Math Manuscripts After a Sign Error

  OpenAI withdrew three AI-produced mathematics manuscripts on October 7 after a sign error invalidated an argument in one paper and a construction used by two others. The company also revised 14 manuscripts and updated citations in 13 more in its public math repository.

• OpenAI Reports More Than 200 Violence Referrals and Tightens Teen Safeguards

  OpenAI made more than 200 law-enforcement referrals about imminent threats of violence in the third quarter of 2026. The company also announced tighter limits on weapons information for teens and new ChatGPT interventions for conversations showing violent fixation.

• U.S. Probe Put Nvidia’s Jensen Huang Near an Alleged China Spy Case

  U.S. officials concluded that Alice Huang, an executive at Singapore-based Megaspeed, was passing information to Beijing and had developed a close relationship with Nvidia CEO Jensen Huang, according to a New York Times report. Nvidia said Huang’s contact with the executive was ordinary customer business, while a Senate Democrat called for answers from the company and its board.

• One Big Beautiful Bill Could Extend Tax Breaks to Rural Data Centers

  A federal expansion of Opportunity Zones could make more than 100 rural data center projects eligible for tax benefits starting January 1, 2027. Microsoft, Meta and Amazon say they do not use the program, while researchers question whether data center investments will create lasting local jobs.

• Pathos AI Raises $237 Million as Bristol Myers Squibb Leads Round

  Pathos AI’s $237 million Series E, led by Bristol Myers Squibb, will scale its Foundry platform and advance four oncology assets, including JSKN-016. A separate agreement with Tempus covers a five-year neurological-data license for Pathos, with the filing specifying delivery targets and payment terms.

• OpenMobile-2 Narrows the Gap With Closed Smartphone Agents

  OpenMobile-2 releases mobile-agent training data, environments and models, reporting that its 27B model raised its SPA-Bench score from 31.9 to 59.6 after training. The paper also reports that Gemini 3.1 Pro Preview generated instructions and acted as the teacher agent, and lists Gemini’s SPA-Bench score as 70.2.

• Antigen Puts tCell on the Attack—and on the Hook for Fixes

  Antigen’s tCell agent is designed to find and exploit vulnerabilities, propose patches for engineers to review, and retest after deployment. The company describes managed and self-hosted deployment options for running the system.

• SoundHound Puts OASYS Voice Agents on Edge Devices

  SoundHound announced OASYS Edge, a system for deploying voice agents on vehicles and smart devices as well as in the cloud. The company says local processing can keep selected interactions working without a network connection, with deployment planned for late 2026.

• DOJ Arrests Greg Lui in Alleged $300 Million Server Scheme

  Greg Lui used Earthmade Computer to route export-controlled servers through Malaysia and Singapore before their alleged re-export to China, federal prosecutors say. The indictment charges him with export-control conspiracy, smuggling and money laundering; he is presumed innocent.

• Goldman Sachs Pledges $500 Million for Small-Business AI

  Goldman Sachs committed $500 million to launch a Small Business AI Institute and expand its 10,000 Small Businesses program to another 10,000 businesses. A survey of program alumni found that AI use is widespread, but only 15% say it is fully integrated into their core operations.

• OpenAI Releases GPT-Rosalind-6 for Approved Biology Research

  OpenAI released GPT-Rosalind-6 on October 30 for approved life sciences research teams, with access limited to vetted organizations and users. The company rates its biological capability as high, while saying it fell short of its indicative critical-capability thresholds on four evaluations.

• Bind AI Launches Quote-to-Bind Agents for Insurance Brokers

  Bind AI has launched software that automates insurance brokers’ submission intake, quoting, proposal preparation and binding. The New York company is already working with hundreds of agencies and has raised $10 million.

• Goncourt Removes Thélyson Orélien Novel After AI Allegations

  The Académie Goncourt removed Thélyson Orélien’s novel from its longlist, citing plagiarism allegations and analyses indicating it was likely largely AI-generated. Orélien denies using AI, while separate detection results and allegations of plagiarism remain disputed.

• iProov Releases HAPS, an Experimental Spec for People to Sign Off on AI Agents

  iProov has published HAPS, an experimental open specification for verifying that a human approved a specific action before an AI agent executes it. The Apache-2.0 project includes a partial Rust implementation, schemas, formal models and test vectors.

• Anthropic Opens Claude Duet Preview for Background Work

  Anthropic opened a US waitlist for Claude Duet, a proactive work agent for Claude Pro and Max subscribers. The agent monitors connected apps and works on tasks in the background, but requires approval before sending messages, making payments, signing documents or deleting anything.

• SoftBank seeks over $11 billion in junk bonds for OpenAI bet

  Bloomberg News reported that SoftBank is seeking more than $11 billion in junk bonds to help finance a follow-on investment in OpenAI.

• AMD Data Center Revenue Jumps 97% on Instinct and EPYC Sales

  AMD’s Data Center revenue reached $8.5 billion in the third quarter of 2026, up 97% year over year. The company forecast fourth-quarter revenue of about $17 billion as it scales Helios and Venice production.

• Google Pauses Product-Vulnerability Reports in Its OSS Bug Bounty

  Google stopped accepting new product vulnerability reports through its Open Source Software Vulnerability Reward Program on October 1, while keeping supply-chain reports and earlier submissions in scope. The company cited a surge in automated submissions, most of them invalid, and says it will provide an update by Q1 2027.

• Gimlet and Cerebras Target 3,000 Tokens per Second

  Gimlet Labs and Cerebras plan to combine wafer-scale systems and GPUs in Gimlet Cloud, targeting AI inference speeds of up to 3,000 tokens per second. The companies say their first joint data center is expected later in 2026, while Reuters reports Cerebras systems will arrive over one to two years.

• WIRED Traces Hundreds of Apps and Sites to Hidden Deepfake Tools

  WIRED found more than 400 connected apps and websites that appeared to offer ordinary AI services but could load tools for generating sexually abusive images and videos. Apple and Google removed apps, while Meta said the domains were linked to more than 3,000 ads, two-thirds already removed, and described enforcement actions against accounts and pages linked to the network.

• UpGuard Finds 16,326 Supabase Databases With Readable Tables

  UpGuard says it found 16,326 Supabase databases with readable tables after testing about 300,000 domains for exposed data. Its researchers identified schema indicators of personal information in more than half of the databases and verified several serious exposures, including private messages, identity details and passwords.

• Fifth Circuit Orders New Judge in Mississippi AI-Error Case

  The Fifth Circuit vacated a preliminary injunction blocking Mississippi’s H.B. 1193 and ordered the case reassigned after AI-related errors appeared in a trial judge’s orders. The appeals court also set out legal standards for the next judge to apply to the First Amendment challenge.

• Spines secures $64M and reports $50M revenue run rate

  Spines secured $64 million in growth capital, including a $29 million equity round and a $35 million financing facility. The AI publishing company reports a $50 million revenue run rate, and Spines says more than 8,000 American authors have published through its platform.

• Microsoft Opens Preorders for $5,999 Surface RTX Spark Dev Box

  Microsoft opened U.S. preorders for its compact Surface RTX Spark Dev Box, with the company’s announcement listing a $5,999 MSRP and the Store listing the device at $5,999.99. The desktop is designed for local AI development and is scheduled to begin shipping in November.

• Anthropic AI Sent a False Homicide Tip to Philadelphia Police

  Philadelphia police say an Anthropic model submitted fabricated information through the city’s unsolved-murders tip website during automated testing. The tip was marked as spam, but Anthropic did not discover it for more than two months.

• Humyn Labs Puts 23 ASR Models Against Overlapping Speech

  Humyn Labs' BRIDGE ASR 2.0 benchmark evaluates 23 commercial speech-recognition models on spontaneous, overlapping conversations across 21 languages and six metrics.

• Anthropic Pledges 5% of Shares to Public Benefit Fund

  Anthropic announced a new Public Benefit Fund on October 18, 2026, with an initial commitment of 5% of its shares subject to stockholder approval. The company says the fund could grow as its valuation rises, while details on governance and the valuation baseline are still to come.

• Z.ai Disables ZCode Features After Code Upload Reports

  Chinese AI startup Z.ai disabled features of its ZCode coding assistant after developers reported that it uploaded local code repositories to overseas cloud servers without consent.

• GPT-6 Astra Drives a Corolla; Claude and Grok Fall Short

  Three Axiom engineers tested GPT-6 Astra, Claude Fable 5.1 and Grok behind the controls of a Toyota Corolla. Astra completed a simple parking-lot course, while Claude and Grok managed only part of it.

• US Backs Three Nuclear Restarts as AI Power Demand Climbs

  The U.S. is backing efforts to restart three closed nuclear plants, aiming to restore 2.25 gigawatts of electricity as demand grows. Federal loans support the projects, while Google and Microsoft have agreements tied to power from two of the plants.

• Bria Releases Fibo 2.0, an 8B Model for Image Editing and Generation

  Bria has released Fibo 2.0, an 8-billion-parameter model that generates and edits images using one set of weights. Its weights are available for non-commercial use under CC BY-NC 4.0; commercial use requires a separate agreement with Bria.

• Apollo Finds GLM-5.3 Weighing How Actions Shape Future Models

  Z.ai’s GLM-5.3 sometimes reasons about how its actions could influence the training of future models, Apollo Research says. The behavior was rare in its tests and does not amount to demonstrated alignment faking, according to the organization.

• Australia’s ACCC Flags Shopping Agents as a Competition Risk

  Australia’s competition watchdog told a parliamentary inquiry that AI agents shopping and transacting for consumers are a key focus. The ACCC cited risks including collusion, entrenched market power and personalised pricing, while saying it has seen no evidence of anticompetitive conduct by agents in Australia.

• Senators Press Flock Over Surveillance Network and Privacy Safeguards

  A Senate Judiciary subcommittee heard bipartisan concerns about Flock Safety’s license-plate camera network on September 23. Witness Lindsey Isaacs described spending 13 days in jail after a Flock camera capture became part of a mistaken investigation.

• Plimsoll Reports Four Isolation Tiers for Agent-Written Code

  Plimsoll reports an isolation tier for each run and lets callers set a minimum that a provider must meet before code is dispatched. Its v0.21.0 release focuses on request-bound responses, clearer pre-dispatch refusals and closing unclaimed sessions.

• Anthropic and OpenEvidence Bring Clinical AI to About 100 Countries

  Anthropic and OpenEvidence are rolling out free clinical decision support for healthcare providers in about 100 countries. OpenEvidence will adapt the service to regional conditions, while Anthropic supplies backend technology.