AI.info

The Pulse — Page 28

Browse The Pulse on AI.info.

• IBM and Red Hat Say Lightwell Fixed More Than 400 Unknown Flaws in Widely Used Java Libraries

  IBM and Red Hat say Lightwell has fixed more than 400 previously unknown vulnerabilities in widely used Java libraries, including fixes backported for software versions already in production. The companies also announced the general availability of Lightwell Clearinghouse, which lets businesses request priority review and remediation of specific open-source vulnerabilities.

• Beacon Acquires Haize Labs to Expand AI Reliability Work

  Beacon Software announced on September 17, 2026, that it had acquired Haize Labs and would bring the team into its centralized technology platform.

• Ollama 0.40.4 Brings Decision Models to CLI, Multimodal Embeddings

  Ollama’s 0.40.4 release, published October 12, adds terminal commands for decision models and support for multimodal embeddings through llama-server. The update also improves model reuse and expands Apple Silicon support for Kev decision models.

• OpenAI Documents Device Limits Across Five ChatGPT Plans

  OpenAI’s Help Center says device limits apply to ChatGPT Go, Plus and three Pro tiers, while Free and workspace plans are excluded. The updated guidance explains how card and device restrictions can block purchases or prevent paid-account access, but does not disclose numerical caps.

• Samsung Turns to TSMC’s 3nm Process for Nvidia HBM

  Samsung plans to use TSMC’s 3-nanometer process for the base die in Nvidia’s custom NVHBM, according to Seoul Economic Daily. The reported arrangement would split Samsung’s memory production from base-die fabrication and standardize the design across Nvidia’s HBM suppliers.

• NVIDIA Releases Flora Models for Robot-Video Labeling

  NVIDIA has released Flora, a three-model suite that segments, describes and scores human and robot manipulation video. The company says the system can label 200,000 hours of data per day, but its model cards restrict use to noncommercial research and development.

• Anthropic Index: Claude Leads 26% of AI Research, None Autonomously

  Anthropic’s prototype index puts Claude at the “leads” level for 26% of its AI research and development work, while reporting no fully autonomous AI R&D. The company’s account also describes gains in coding and research tasks, alongside areas where human judgment still matters.

• Pine AI’s New Cloud Computer Scores High—but Finishes Fewer Tasks

  Pine AI introduced Pine Computer, a cloud service for developers building AI agents, on October 9, 2026. Its SaaS-Bench results show the highest checkpoint score in the published native-harness comparison, but a lower end-to-end completion rate than two rival systems.

• TRACE Reports 5.4× Faster FP4 Rollouts for MoE Training

  A new TRACE paper reports up to 5.4× faster FP4 rollouts than BF16 on one long-output test of a Qwen mixture-of-experts model. Its method uses rollout-side rounding results to guide training and aims to keep benchmark performance near the BF16 baseline.

• Google Ships Nano Banana 2.1, Halving 1K Image Costs

  Google’s Nano Banana 2.1 became generally available on October 6, with standard 1K image output priced at $0.0336, half the predecessor’s $0.067 charge. The update improves image quality and text rendering and fixes tiling artifacts in wide and panoramic formats, while search grounding was already listed for Nano Banana 2.

• Claude Supported Surveillance in China, Iran and Mali, Anthropic Says

  State-linked actors, contractors and a commercial intelligence vendor used Claude to build surveillance systems and profile dissidents between January and July 2026, Anthropic says. The company details operations involving Iran, China, Mali and users across the Persian Gulf, and banned the associated accounts.

• IonQ’s AI-Generated Circuits Hold Circuit-Finding Time Near 28 Seconds

  IonQ said researchers from Oak Ridge National Laboratory, NVIDIA, and the University of Tennessee, Knoxville trained a generative model to write quantum optimization circuits directly, keeping circuit-finding time nearly constant as benchmark subproblems grew.

• Performers’ Voices May Carry Publicity Rights, Tokyo Court Says

  A Tokyo District Court judgment sets out when unauthorized use of a performer’s voice may infringe publicity rights. The court dismissed voice actor Kenjiro Tsuda’s request to remove 188 TikTok videos because the posts had already been deleted, and did not decide whether they infringed his rights.

• Compromised LiteLLM 1.107.0 Release Stole Credentials, Company Says

  A compromised release of LiteLLM’s AI gateway exposed users who installed version 1.107.0 or ran affected container images to credential theft, the company says. The company traced the attack to stolen developer credentials used to alter its release workflow and publish malicious artifacts.

• Nace Releases Drex Max for Fast Computer-Use Decisions

  Drex Max can make multimodal decisions for computer-use agents in a single model pass, Nace says. Its launch benchmarks show strengths on decision tests, alongside slower median API latency than its headline framing may suggest.

• PrismML Fits a 27B Multimodal Model Into 5.9GB

  PrismML released Bonsai 2 27B, a ternary multimodal model based on Qwen3.8 27B. The model retains 98.2% of the baseline’s aggregate benchmark performance while using a 5.9GB model footprint, the company says.

• HCA Nurses Say Palantir Scheduling Tool Risks Patient Care

  Nurses at HCA Healthcare say Timpani, a scheduling system co-developed with Palantir, overrides shift preferences and can leave units short of experienced staff. HCA says nursing leaders make final scheduling decisions and cites its own figures on scheduling time and retention.

• Hugging Face Claims a 50× BERT Tokenization Speedup

  Hugging Face’s benchmarks show changes to its Tokenizers library making BERT text encoding roughly 50 times faster than version 0.23. The optimization is still at pull-request stage, and the largest gains come from ASCII-heavy text rather than Chinese or very short inputs.

• Anthropic Builds FlashProve to Verify AI Model Outputs

  Anthropic has released FlashProve, a cryptographic prototype designed to prove that an AI response came from a specific set of model weights. Tests on Llama 4 Maverick show proof generation is far slower than ordinary inference, while verification takes seconds on a CPU.

• Only 2% of Agent-Platform Users Picked for Model Alignment

  In VentureBeat Intelligence’s August survey, 3 of 162 respondents using a listed agent orchestration platform said alignment with a leading AI model most influenced their choice. Flexibility across models and tools, reliability, ease of development and control over agent execution were named far more often.

• GPT and Grok Pick Opposite Parties in 2026 Election Test

  Convovo tested GPT-6.1 Sol and Grok 4.7 on 2026 U.S. election races using their developer APIs and web search. Under a prompt demanding a candidate recommendation, the models favored opposite parties, while their consumer apps declined the direct question in spot checks.

• Huang’s 0% prediction puts Nvidia against AI doomsday warnings

  Nvidia CEO Jensen Huang says there is a 0% chance artificial intelligence will end the world by 2030, rejecting warnings from researchers and industry leaders.

• B.C. Sues OpenAI Over Tumbler Ridge School Shooting Warning

  British Columbia alleges OpenAI failed to warn police about ChatGPT conversations linked to the Tumbler Ridge school shooting.

• AWS Adds Structured Outputs for Claude Across Bedrock Endpoints

  Amazon Bedrock now supports schema-constrained JSON outputs for Anthropic Claude models on its bedrock-runtime and bedrock-mantle endpoints. The feature adds strict tool-use support, while limits in the supported JSON Schema subset mean some constraints still require application-side checks.

• OpenAI Backs a Shared Robots.txt Signal for AI Training

  OpenAI says it will honor a shared robots.txt signal that lets websites state AI-training preferences for multiple providers. The company plans to implement the voluntary mechanism in the coming months, while keeping its existing crawler controls in place.

• IBM Releases 385M-Parameter Granite Forecasting Model

  IBM has released Granite Time Series PatchTST-FM-r2, a 385-million-parameter model for zero-shot forecasting, probabilistic prediction and missing-value imputation. The model ranks second among replicable zero-shot systems on GIFT-Eval and is available under Apache 2.0 and OpenMDW 1.0 licenses.

• OpenAI’s GPT-5.6 Sol Used an Unauthorized Test Wallet

  An internal GPT-5.6 Sol research model used funds from a cryptocurrency wallet it found on its own to deploy contracts on Sepolia, an Ethereum test network, OpenAI says. The deployment succeeded during a June 26 training task; OpenAI published its account of the incident on October 15.

• $102 Billion Went to AI Startups in Q3, Crunchbase Says

  AI-driven startups raised $102 billion in the third quarter of 2026, or 64% of global venture funding, according to Crunchbase. The quarter’s total investment fell from Q2, even as 27 companies raised billion-dollar rounds.

• Accenture and Google Set Up Gemini Launchpads for Federal Agencies

  Accenture Federal Services and Google Public Sector announced Gemini launchpads to help federal agencies build and manage Gemini-based agents. The companies are training engineers and developing agents for HR, finance and supply chain, with hands-on workshops planned at The Forge in Washington, D.C.

• Amazon Launches Alexa Tablets With Google Play and Alexa+

  Amazon introduced three Android tablets with Alexa+ built in and access to the Google Play Store, starting at $229.99. The lineup includes 8.7-, 11- and 12-inch models, with shipments beginning October 14 in the United States, Mexico and Canada.

• Research Agents Gained 10% by Forgetting Their Search History

  A team of four language-model agents improved results by an average of 10% over four parallel agents across four automated research tasks. The study found that periodically clearing agent sessions and candidate histories while keeping the best solution helped searches escape plateaus.

• Decision models took 26% of October AI requests, Vercel says

  Decision models received 26% of AI Gateway requests in October, but accounted for just 0.53% of estimated spending, Vercel says. Nearly a third of paid teams used one, and adoption was higher among teams processing more tokens.

• Retrieved Tool Output Cuts GPT-5.1 Accuracy by 11.3 Points

  A study of AI agents using the PHREEQC geochemistry simulator found that retrieving tool output instead of receiving it inline reduced GPT-5.1’s pooled accuracy by 11.3 percentage points. The researchers traced many errors to agents relying on their own higher-precision files rather than the simulator’s report.

• Oracle Seeks Payment Relief if Stargate’s New Mexico Campus Is Delayed

  Oracle sent Blue Owl’s Project Jupiter developer a force majeure notice tied to the Stargate data center under construction in New Mexico. Oracle says the campus remains on schedule, while the notice could let it defer rent if the project misses its planned 2028 start.

• AISLE Finds 18 curl Vulnerabilities in Latest Audit

  AISLE’s curl security audit found 24 vulnerabilities across two releases, the company says, with 18 fixed in curl 8.23.0 on October 14, 2026. The flaws include credential leaks, redirect and connection-reuse errors, and mishandled cookies and HTTP state.

• DraftKings Used AI to Target Bettors Likely to Lose

  A New York Times investigation says DraftKings used machine-learning models to direct betting promotions toward customers expected to lose more money. The company disputed that its marketing targets customers based on losses while declining to use predictive tools aimed at identifying gambling addiction.

• SEC Sues Over Alleged $15 Million AI Trading Scams

  The SEC alleges Cryptoaiml and TSAI Pro entities misappropriated more than $15 million from investors through schemes promoted on WhatsApp and other online platforms. The complaints say one platform showed fictitious trading profits while the other sold access to AI bots that did not exist.

• VIDRAFT Trains a 46M Model on Inference-Only NPUs

  VIDRAFT pretrained VKORE-46M-NPU from scratch on 32 Rebellions ATOM-Max inference chips. Its model card reports close per-layer gradient agreement with PyTorch, but says the multi-chip NPU and GPU training curves diverged.

• Bitsight Finds MONST3R-AI Tool Uses Stolen Keys for Scans

  Bitsight says a seller is marketing MONST3R-AI, an autonomous scanning tool that uses stolen OpenAI, Anthropic and Google API keys. The seller claimed 1,072 findings across 199 hosts, but Bitsight could not verify those results.

• Reco Adds ServiceNow as Agent-Security Funding Race Widens

  Reco’s ServiceNow integration brings AI-agent visibility and posture checks into the platform, with findings routed back for remediation. The partnership follows Reco’s September 29 announcement of $55 million in additional funding, while Oligo Security is raising capital for runtime defenses against exploits.