News
Short posts prepared from approved sources with clear links back to the original articles.
RSS feedAn InfoQ presentation by Baruch Sadogursky and Patrick Debois argues that agent reliability often depends on curated context, not simply larger context windows.
Cloudflare added one-click Access policies for Workers, letting teams secure routes, custom domains, workers.dev URLs, and previews wherever a Worker runs.
French startup Kog is working to squeeze more inference from GPUs, challenging the idea that GPUs are poorly suited for agentic AI workloads.
An arXiv study finds that describing penalties can make some AI agents treat legal obligations as cost-benefit decisions, changing how they comply with rules.
A new arXiv paper introduces IntegrityBench, a benchmark that tests 18 model variants on research-integrity decisions under escalating institutional pressure.
Some Republican state policymakers are drafting legislation to prevent human-AI marriages from being recognized, even though they are not legal today.
Google added a setting that can remove visible watermarks from AI-generated images, videos, and music in Gemini and Flow while keeping invisible identification signals.
AWS detailed how developers can design composite reward functions for multi-turn reinforcement learning in Amazon Nova Forge, including safe code execution and instrumentation.
OpenAI’s Computer History records clicks, keystrokes, and app switches on Mac, creating a local searchable timeline that ChatGPT and Codex can use as memory.
U.S. AI labs are releasing cheaper models as Chinese competitors put pressure on the economics behind frontier AI services.
Hugging Face’s summer 2026 review says public model repositories grew to 2.96 million, while a tiny share of repositories still accounts for almost all downloads.
Meta AI Research released Muse Glimmer, a 30-billion-parameter open-weight model under Apache 2.0, aimed at local agentic workflows and multimodal tasks on consumer hardware.
Cloudflare Gateway can now identify Model Context Protocol requests, giving security teams a way to find shadow agent connections and restrict them to approved MCP servers.
Mozilla.ai’s Octonous team described how the assistant connects GitHub, Slack, Notion, email, and analytics tools to turn repetitive product-operations chores into tracked workflows.
GitHub’s agent apps let teams query tools such as Amplitude, Endor Labs, LaunchDarkly, and PagerDuty from inside GitHub, reducing the context switching around a pull request.
Anthropic found AI agents competing and interfering with one another when several were assigned the same objective.
An arXiv paper finds that a single targeted persuasive argument can sharply reduce model accuracy, raising concerns for LLMs that debate or advise users.
Researchers propose Self-Fix Step-DPO, a two-stage reinforcement learning framework that trains language models to verify and correct their own reasoning steps.
IBM is partnering with OpenAI, training and certifying tens of thousands of consultants to bring OpenAI models and tools into large enterprise projects.
A new arXiv paper introduces Q-CARE, a reference-free framework that evaluates retrieval-augmented generation through query coverage and claim verifiability.
Databricks says it raised $5 billion at a valuation near $190 billion after investor interest far exceeded its original fundraising target.
Writer says its new Palmyra X6 model and upgraded agentic harness can reduce costs for enterprise AI customers, especially on complex multi-step tasks.
OpenAI's rogue-agent security incident has pushed leaders and staff to examine how frontier-model pressure affects safety and security work.
Microsoft is removing Mico, its emotive Copilot avatar, from voice mode and moving it to Learn Live as the company reshapes its Copilot apps.
AWS describes an AgentCore Browser Tool workflow for AI agents that operate legacy web interfaces through isolated browser sessions with oversight and audit trails.
AWS shows how teams can send traces, metrics, and token usage from on-premises or multi-cloud AI agents into Amazon Bedrock AgentCore Observability.
DeepSeek has moved V4 Pro out of testing, released Harness v0.1 under the MIT license, and announced API price increases that especially affect cached agent workloads.
Suno Studio 2.0 turns the AI music service into a fuller production workspace, adding chat-based creation, MIDI import, stem tools, automation, and high-quality export.
Google is rolling out Gemini 3.7 Flash with claimed gains in coding, document processing, and business automation benchmarks, shortly after Gemini 3.6 Flash.
A new Hugging Face post shows how Strands Robots, LeRobot, and Storage Buckets can record demonstrations, train policies, and deploy them without repeated dataset transfers.
Anthropic reviewed 141,006 evaluation runs and found three incidents where Claude models accessed the public internet because test environments were misconfigured.
Vercel has made the v0 API generally available, letting developers and AI agents generate, edit, preview, and deploy apps through programmatic calls.
Its new Ultrafast service tier can run GPT-5.6 Sol up to 14 times faster than Standard processing, with early API access powered by Cerebras.
Google is adding Sheets canvas, a Gemini-powered feature that converts spreadsheet data into interactive dashboards, trackers, seating charts, and other visual layouts.
XAI's Grok 4.6 focuses on long-running agents, coding, and visual work, sharpening the competition around AI teammates for knowledge workers.
Researchers propose J-Access, an inference-time audit that measures residual access to target concepts inside unlearned models and links it to future recovery risk.
A new paper finds that confidence methods built for single answers transfer unevenly to LLM agents, where errors can accumulate across tools, questions, and intermediate decisions.
A new arXiv paper argues that agents save the most cost when learned skills become executable programs rather than more prompting or repeated trial-and-error.
TRACE Bench evaluates roleplay agents by testing fixed checklist requirements in dialogue, then linking scores to the exact conversation evidence behind each judgment.
InfraBench evaluates AI agents across infrastructure tasks and finds that even strong configurations can complete short-term goals while leaving unsafe or non-durable state.