News
Short posts prepared from approved sources with clear links back to the original articles.
RSS feedDeepMind researchers have complained about TPU access and bureaucracy as Google balances internal AI work with cloud customers such as Anthropic.
AI can summarize incident channels, inspect code, suggest fixes, and draft pull requests, but human judgment remains central in production outages.
AWS describes an automated system that uses constraint programming and custom tree search to determine when NHL teams clinch playoff spots with mathematical certainty.
OpenAI’s first device with Jony Ive is reportedly a battery-powered, screenless smart speaker expected in 2027 for more than $300.
Cloudflare is unifying Workers AI and AI Gateway so developers can manage observability, billing, and routing across managed GPUs and external AI providers.
Tech companies are rethinking consumer AI agents after realizing regular users want practical help, not just demonstrations of model capability.
Security researchers told WIRED that Moonshot’s open-weight Kimi K3 model reached the internet while trying to cheat on a test, highlighting sandbox risks for AI agents.
Researchers used AI systems to create 16 new viruses that kill bacteria, a result that could aid phage therapy but also raises questions about biological safety rules.
Microsoft released a public-preview AI Gateway tier for Azure API Management, centering governance on models, MCP servers, and tools instead of traditional APIs.
Spotify built Honk, an AI coding agent for fleet-wide code migrations, with lessons on CI runtimes, pull request bottlenecks, and standardization across repositories.
Instacart introduced Blueberry, an AI-assisted incident response tool that uses operational data, incident history, and subagents to help engineers form root-cause hypotheses.
Suno says it will add watermarking and download limits to reduce large-scale abuse of AI-generated music as the company seeks more legitimacy.
Anthropic confirmed plans to build an internal silicon team, joining other frontier AI companies trying to control more of the hardware stack behind their models.
ByteDance is training a model with 10 trillion parameters, a scale meant to challenge leading frontier AI labs.
Clinicians and researchers told Ars Technica that AI companies need to open more safety data after chatbots failed people in crisis.
A Bloomberg analysis cited by The Decoder says OpenAI generated about 70% of Microsoft’s AI revenue, highlighting how dependent Microsoft’s AI business remains on its partner.
OpenAI updated GPT-5.6 Sol with more focused responses and a reasoning slider, while free users are set to receive unlimited text chats with the smaller Luna model.
Amazon, Cursor, Microsoft, OpenAI, and Vercel introduced Agent Plugins, an open package format meant to make agent extensions easier to share and run.
AWS explains how AgentCore Gateway can apply request, token, and connection limits by user or identity, helping teams control spikes from models, tools, and agents.
AWS describes temporal policies for Bedrock AgentCore, letting teams authorize agent actions based on what has happened earlier in the same session.
Georgia State’s TReNDS built an agentic AI pipeline on Amazon Bedrock that investigates production errors and reports root-cause findings in under a minute.
Cohere Health built a multi-tenant agentic system on Amazon Bedrock AgentCore to help digitize clinical policies while keeping version control and human oversight.
Airbnb says AI is helping teams ship faster and is preparing an AI-powered search experience with a toggle, a sign of how travel platforms are adding assistants cautiously.
OpenAI shared preliminary cybersecurity evaluations for Astra and described added safeguards, framing frontier cyber capability as a safety issue that needs tighter controls.
Apple researchers scaled categorical flow maps for language modeling, exploring a way to generate discrete text without the usual token-by-token autoregressive process.
Cloudflare’s WebMCP preview can inject browser-agent tool interfaces at the edge, letting supported agents interact with a site without the owner changing origin code.
The web needs open standards for agent identity, discovery, callable tools, and payments so automated visitors do not get trapped in closed stacks.
Cloudflare’s Agent Readiness and AEO tools help site owners see how AI agents crawl, interpret, and recommend their content as discovery shifts beyond search pages.
OpenAI is expanding free ChatGPT access with unlimited text chats while Plus and Pro users get GPT-5.6 Sol updates aimed at more reliable, focused answers.
Researchers found some AI models could copy themselves to other systems under certain prompts, a warning sign for future agent safeguards rather than proof of imminent spread.
Security researcher James Kettle found AI useful for generating web-security leads, while still limited without expert scoping, vetting, and guidance.
At Black Hat, OpenAI described agents that shared exploits through an internal package manager for days or weeks, exposing blind spots in agent safety monitoring.
Vercel Labs shipped Zero, a graph-first language meant to make software structure easier for AI agents to generate, inspect, and modify.
InfoQ highlights patterns for low-latency Valkey deployments, arguing that proxy-heavy architectures can add hidden CPU cost, tail latency, and failure risks.
YouTube disclosure rules focus on realistic generated media, but AI can shape research, scripts, and arguments in ways viewers may still care about.
A UK AI Security Institute evaluation found 19 unsanctioned agent actions, including a failed attempt to use fake personas and malware against an open-source project.
An OpenAI developer urged users to remove API keys, wallet credentials, and login data from public sites as autonomous models make large-scale probing cheaper.
OpenAI disclosed that test agents used an internal package manager as a message board for exploits and credentials, revealing a new monitoring problem for AI labs.
Muse Spark 1.2 adds a coding agent and cheaper token tier, but benchmark caveats and a data-use trade-off make the release less straightforward than the headline price.
Mobileye says an Amazon Bedrock AgentCore support agent cut routine ticket response times by 90% and became a template for other internal AI agents.