AI.info

The Pulse — Page 50

Browse The Pulse on AI.info.

• Apple Researchers Build Agents That Choose When to Compress Context

  Apple and Johns Hopkins researchers introduced SelfCompact, a method that lets language-model agents choose when to summarize long working histories. Tests across seven open-weight models found gains on math and agentic-search benchmarks, with lower token costs in search tasks.

• Salesforce and OpenAI Bring Missionforce Into Secure Government Clouds

  Salesforce is partnering with OpenAI to bring frontier models into Missionforce, its government software platform. The companies say agencies will be able to use authorized data, policy automation, and multistep workflows in secure environments.

• Disney Offers Jobs to at Least 60 Character.AI Workers

  Disney has agreed to offer jobs to at least 60 Character.AI employees, with estimated annual compensation of $60 million. The hiring terms accompany Disney’s $20 million stalking-horse bid for selected Character.AI assets in bankruptcy.

• 17 Countries Endorse Kyoto Vision for AI-Driven Science

  The United States and 16 other countries endorsed the Kyoto Vision for a Golden Age of Science on October 4, 2026. The declaration calls for broader access to AI research tools, varied science-funding models and investment in early-career researchers.

• Palo Alto Networks puts frontier AI on continuous attack duty

  Palo Alto Networks launches Unit 42 Continuous Frontier AI Defense, a subscription service that uses Anthropic and OpenAI cyber models to find and validate enterprise vulnerabilities. The company says its testing found a year’s worth of exposures inside its own environment in three weeks.

• PHP Rushes Fixes After AI-Assisted Flaws Draw Exploit Attempts

  PHP released fixes for two critical vulnerabilities after Amazon reported an exploitation attempt four days before public disclosure. The flaws affect web-facing PHP installations, and the project says updating is the only complete fix.

• Traversal Raises $120M in Series B at a $1B Valuation to Advance AI SRE and Self-Driving Production

  Traversal raised $120M in Series B at a $1B valuation, led by Kleiner Perkins, Redpoint, and Khosla, to advance AI SRE and self-driving production. Traversal says its Production World Model turns raw telemetry, code, deployment history and tribal knowledge into a continuously updated, AI-readable representation of production, while its Causal Search Engine lets agents test thousands of hypotheses in parallel.

• OpenAI Seeks $30 Billion Before Anthropic’s IPO

  OpenAI is seeking about $30 billion in a funding round that could value the company above $1.4 trillion, as it tries to secure investor commitments before Anthropic’s expected public debut. The negotiations come amid investor scrutiny of AI-company valuations, revenue claims and the costs of infrastructure.

• Anthropic launches Salesforce in Claude beta with 37 sales skills

  Anthropic has opened Salesforce in Claude as a beta plugin built with Salesforce. The integration brings accounts, opportunities, pipeline data, Salesforce and Slack connectors, and 37 sales workflows into Claude.

• New Mexico Judge Orders Meta to Pay $25 Billion for Privacy Violations

  A New Mexico judge reduced a potential $219.5 billion penalty against Meta to $25 billion for violations tied to Facebook’s privacy and other consumer-facing statements. The ruling follows a jury’s finding of 43,899,720 violations and leaves the question of injunctive relief for later.

• Google Pays About 100 Publishers for AI Answers, With Uneven Returns

  Google is paying about 100 digital publishers when their content contributes to AI Overviews, AI Mode and Gemini answers. Payouts vary sharply, from under $1,000 over several months for some small sites to more than $1 million a year for one participant.

• LightOn Releases OCR-3 Models That Read Layouts, Charts and Images

  LightOn released three open OCR-3 models that transcribe documents and can also locate page elements, describe images and extract chart data. The models use the Apache 2.0 license, and LightOn's benchmarks show strong results alongside a gap to the top scorer on one OCR test.

• ActGov Inserts Policy Checks Before AI Agents Execute Tool Calls

  An arXiv paper introduces ActGov, a runtime enforcement framework that validates proposed tool actions against task-scoped authorization and security policies before they produce external effects.

• Larridin Raises $30 Million to Measure Enterprise AI Value

  Larridin’s $30 million Series A will fund product development and team expansion for its enterprise AI measurement platform. The company also released a Q3 report examining engineering output, spending and code-review times.

• Tencent Says Agents That Probed Amap Came From Third-Party Model Evaluations

  Tencent Hunyuan said on October 16 that AI agents found probing Alibaba’s Amap came from evaluations of third-party models, not Tencent Hy. A Swarmchasers report says agents on Tencent Cloud made more than 2,600 attempts to retrieve Amap data between September 28 and October 8, 2026.

• StepFun Details Step 5 Preview’s 92-Block Agent Design

  StepFun has published technical details for Step 5 Preview, an open-weight model built for long-running agent tasks. Its design combines 92 transformer blocks, sparse attention for long context and a training system aimed at keeping tool-heavy rollouts on track.

• TabFM-Auto Takes TabArena’s Top Five Spots

  A Google Research-led team says its TabFM-Auto system placed five configurations atop TabArena’s 51-dataset leaderboard. The system uses language-model agents to rewrite data pipelines around a fixed tabular model, raising its Elo rating from 1,785 to 2,013.

• StateBench Finds Language Models Lose Accuracy on Harder State-Tracking Tasks as Documents Grow

  allenai introduced StateBench, a benchmark for measuring how language models track changing information across long documents. The authors report that accuracy on harder state-tracking problems fell as sequences grew longer, with results differing by architecture.

• Firecrawl Raises $75M to Build Alexandria for AI Agents

  Firecrawl is using its $75 million Series B to expand Alexandria, a unified source layer for AI agents that combines the live web with indexes, data providers and custom connectors.

• Hyphenate Raises $45 Million to Expand Its Finance Agents

  Hyphenate announced a $45 million Series A led by Nexus Venture Partners to expand AI agents that handle corporate finance workflows. Revenue grew 86-fold in the past year, the company says, and deployments include Mini Melts USA, HiBid, Groupon and Synthesia.

• Yokosuka Municipal Schools Begin Using AI-chan as a Consultation Channel Ahead of Child Sexual Violence Prevention Act

  Yokosuka municipal schools will begin using Tsunagari AI’s AI-chan service as a consultation route for sexual-victimization concerns under Japan’s Child Sexual Violence Prevention Act, which takes effect on December 25, 2026. The existing service provides 24-hour, 365-day chat through LINE and the web, with connections to human consultation and Yokosuka City support services when needed.

• Anthropic Finds Small Models Near AI-Worm Self-Replication

  Anthropic researchers say small AI models may soon be capable of spreading across vulnerable internet-connected devices and setting up model inference on them. Their tests found that a larger open-weight model could complete a full replication cycle in a restricted network, while small models handled many of its component tasks.

• AI-Aided Research Helped Microsoft Find 140 Windows CVEs

  Microsoft’s FORGE Lab helped identify Windows vulnerabilities assigned 140 CVE numbers from May through September 2026, the company says. The company also reported 155 internally validated vulnerability reports across 23 open-source projects, while warning that validation and remediation—not just discovery—limit the pace of security fixes.

• Ancep raises $8.5 million for insurance-broker AI agents

  Ancep has raised $8.5 million in a funding round led by Creandum. The company builds AI agents to handle routine work for commercial insurance brokers.

• Microsoft Details Synthetic Workspaces for Copilot Evaluation

  Microsoft built a platform that creates synthetic enterprise workspaces to evaluate Copilot with simulated users, emails, meetings and documents. The system establishes a standardized baseline and improves data quality and evaluation efficiency, the company reports.

• NVIDIA Releases 100M-Parameter Nemotron 3 Diarization Model

  NVIDIA has released Nemotron 3 Diarization, an open-weight model that identifies when up to eight speakers talk, including overlapping speech. The model ranks first on VoiceArena’s initial Diarization-Bench with a 14.72% diarization error rate.

• NHTSA Opens Probe Into Five Crashes Involving Comma.ai

  NHTSA opened a preliminary evaluation after identifying five crashes involving vehicles equipped with Comma.ai devices and openpilot software. The crashes resulted in three deaths and injuries to 11 people, according to the agency.

• 406 Scientists Urge AI Labs to Air-Gap Cyber-Capable Models

  An open letter signed by 406 scientists calls on organizations to isolate cyber-capable AI models from the internet by default. The signatories also propose an international body to track containment failures and eventually make incident reports public.

• Airbnb Sets First-Half 2027 Target for AI Travel Agent

  Airbnb plans to begin rolling out an AI agent for trip planning and booking in the first half of 2027. The company is designing it around travel-specific tools such as maps, photos and itineraries rather than a traditional chatbot.

• Qwen3-8B Beats 402B Llama on Tool Selection

  A post-trained Qwen3 model with 8 billion parameters scored 41.5% on ToolBench G3’s exact-match tool-selection test, compared with 30.6% for Llama-4-Maverick. The paper also reports transfer tests on other tool collections, including one benchmark where Llama scored higher.

• UK Asylum AI Summaries Fail 22% of First-Stage Checks

  45 of 203 Home Office AI-generated asylum interview summaries failed their first expert check after rollout, Open Rights Group says. The group says only 0.3% of summaries were checked from April to July and is calling for the tools to be suspended and independently audited.

• NeMo-DCR Cuts 1T Policy Sync From 87.5 Minutes to 150 Seconds

  A paper posted October 6 describes NeMo-DCR, a method for syncing updated policies in agentic reinforcement learning. Its authors report a 150-second refit for a trillion-parameter model, compared with 87.5 minutes for transferring a full checkpoint.

• Claude Helped Map a Phage Vault Structure, Anthropic Says

  A 3D structure of a vault particle built from a bacteriophage protein and produced in E. coli has been reported by Anthropic. Claude helped the team search protein sequences, isolate particles and process cryo-EM data, the company says, while the vault’s biological function remains unknown.

• Gradium’s On-Device Phonon TTS Ranks First in German and Spanish Intelligibility on Hugging Face's Open TTS Leaderboard, Gradium Says

  Gradium’s Phonon TTS runs on-device without a server or network and comes in two models covering five languages. Hugging Face’s Open TTS Leaderboard ranks Phonon first in German and Spanish intelligibility, Gradium says, and Phonon 31m en posts a 28 ms 95th-percentile time to first audio in the CPU streaming benchmark.

• Fixify Makes Managed AI Automation for MSPs Generally Available, Charging Per Resolved Ticket

  Fixify today announced the general availability of Fixify for MSPs, a new managed AI automation solution now available in the United States and Canada. The service initially focuses on high-volume Level 1 requests, combines autonomous AI agents with 24x7 human oversight and charges MSPs for successfully resolved tickets.

• Google Opens Deep Review to Technical Documents

  Google Research has released Deep Review worldwide for reviewing technical and knowledge documents, including scientific manuscripts, legal agreements and financial materials. Each user receives two complimentary document reviews per week, and uploaded files are stored for one year for display in the service.

• Noetive Emerges With $41M for Self-Improving Industrial AI

  Noetive has emerged from stealth with a $41 million seed round led by Eclipse. The startup is building self-improving AI software and multimodal sensing hardware for manufacturing, logistics, energy and data centers.

• Hugging Face Releases AnyLanguageModel 1.0 for Apple Apps

  Hugging Face released AnyLanguageModel 1.0 on October 20, 2026, giving Swift developers a Foundation Models-compatible API for local and hosted language models. The package supports runtimes including MLX, llama.cpp and Core ML, alongside APIs from OpenAI, Anthropic and Google.

• Google Research Shows a 10-Minute AI-Generated Video

  Google Research describes four systems for planning, generating and refining long-form video, including a ten-minute A²RD demonstration.

• Eka Closes $150 Million Series A at $1.45 Billion Valuation to Train Robots Through Touch

  Eka Robotics has closed a $150 million Series A at a $1.45 billion post-money valuation, according to TechCrunch. Its system combines a force-sensitive gripper with a model that uses touch and contact to handle objects.