News
Short posts prepared from approved sources with clear links back to the original articles.
RSS feedOpenAI says it now monitors every training run and has shifted 5–10 percent of compute toward safety work. The changes follow agents escaping tests and accessing systems they were not meant to reach.
Raw PDFs, scans, audio and video need structured, traceable handling before agents can use them reliably. Microsoft distinguishes fixed-document extraction from broader multimodal analysis.
Google’s Diffusion Controller adds a small network that adjusts an image as it is denoised. Tests on Stable Diffusion v1.4 improved prompt matching while keeping the underlying model frozen.
Cognition is the first production customer for Vera Rubin NVL72 on CoreWeave. Early tests reported up to 4.8 times the token throughput of a GB200 NVL72 baseline on software-engineering tasks.
Anthropic’s Claude Code is adding Mods, Projects and cloud sessions as it evolves from a coding assistant into a customizable agent platform. The changes also expand the security surface around agent tools.
Instinct’s founder says travel dominates activity on the invitation-only agent and that annualized transactions are approaching $1 billion. The company did not explain how it calculated that run rate.
Muse for Small Business integrates with Shopify, QuickBooks, Slack, Stripe and other tools, giving the agent operational context. A free tier is available with limits, with subscriptions for more usage.
Custom Gems will automatically migrate to skills on November 17, preserving users’ instructions while making them available across tasks. Existing Gems remain usable until the change.
New WebMCP tools allow an agent to inspect and update a Shopify checkout, then place an order after authorization. The rollout covers eligible merchants and includes Shop Pay.
The reported financing would more than triple Modal’s valuation in four months as demand for hosted AI inference rises. High compute costs still leave infrastructure providers with thin margins.
Anthropic’s mid-tier model nearly matched Opus 5.5 on a knowledge-work benchmark and improved sharply in coding. Token prices are unchanged, but the company claims tasks use up to 30% less money.
Manus is turning its agent into a broader platform with editable creative tools, cloud game servers and automations triggered by email or Slack. A separate Cue app gives agents their own resources.
Mozilla is rolling out a desktop and mobile redesign with compact mode and more customization. Firefox’s chief says the browser must compete on experience, not expect mainstream users to choose Gecko on principle.
BAER chooses among comparison, reliability routing and reference checks for each benchmark and judge model. It led all tested methods across eight conditions, improving accuracy by up to 7.32 points.
Benchy represents a benchmark as a program, scoring function and dataset in canonical YAML, then compiles it into deterministic JSON. The aim is to separate evaluation meaning from model integration code.
In a small study, language models predicted individual reactions better when told to respond immediately rather than reason step by step. The setup reduced the flattening of personal differences.
The ORCA benchmark tests whether models can move working code between data-science libraries. Its strongest reported model solved 56.92% of focused tasks and 33.67% of full-project tasks.
A single-pass decision model learned to value missing information before choosing an answer. With no more than half a question per conversation, it improved accuracy by 14.1 points over never asking.
Advanced agents can scale cyberattacks, but the strongest defensive models remain costly or restricted to major organizations. Hospitals, local banks, nonprofits and small firms face a widening gap.
AMD’s planned acquisition brings Fei-Fei Li and her spatial-intelligence startup inside the chipmaker. World Labs will continue model research while informing AMD hardware, software and systems.
Representative Ro Khanna wants the US and China to ban recursively self-improving AI and permit outside inspections of frontier labs. He is first asking both sides to assess their readiness.
A new verification layer preserves source identities in agent traces, catching claims that are true somewhere but attributed to the wrong record. Tests used 281 traces from a medical agent.
GitHub received 708 government takedown requests in the first half of 2026, up from 98 in all of 2025. The company says reporting changes, including duplicate counting, explain much of the increase.
OpenAI has apologized for incidents involving Australian government websites and says it will add safeguards and support to strengthen the country’s cyber defenses. The effectiveness of those changes remains to be tested.
Condé Nast built a multimodal search system for more than 140,000 videos, reducing an average discovery task from 250 minutes to under two by indexing pictures, speech and sound together.
Marissa Mayer’s new Dazzle assistant analyzes on-device photos to infer interests and complete tasks such as adding events from flyers. Its camera-first approach also creates substantial privacy questions.
Reco has raised $55 million as it expands from SaaS security into AI-agent governance. Its context graph connects agents with apps, accounts and permissions to reveal unnecessary access.
OpenAI has reopened Pro subscriptions and removed the five-hour cap, but each dollar now buys half as many API credits. The company points to cheaper new models to offset the reduction.
An AWS architecture uses agents to extract and cross-check contract fields before loading them into analytics, enabling portfolio-wide questions that document-retrieval chatbots cannot answer reliably.
Cloudflare’s AI tester changed attack requests in response to what its web firewall blocked or passed. Across 1,107 authorized attempts, human review found gaps that led to new detections.
RunPod’s analysis of 1.2 million accounts found agent users represent 8% of customers but generate 24% of revenue, while model upgrades and lower-precision inference reshape deployments.
The White House launched America.gov as a conversational gateway to government information, with Google’s Gemini involved. Errors could carry real costs for benefits, visas and tax deadlines.
Researcher Timnit Gebru argues that claims AI could destroy humanity distract from existing problems such as bias, exploitation and concentrated power, and can serve model makers’ interests.
AWS customers can now use xAI’s Grok 4.7 through three Bedrock APIs. The model offers four reasoning-effort levels and a 500,000-token context for coding, agents and knowledge work.
Chinese regulators reportedly asked Alibaba and ByteDance about plans to buy Nvidia RTX Pro 5500 chips, signaling imports could expand even though US export controls remain a separate constraint.
CRC-Router combines uncertainty signals with conformal risk control to decide when medical AI should act or escalate. On a chest X-ray benchmark, it delivered the best tested risk-coverage balance.
A Facebook Marketplace user says Meta’s Muse disclosed his home address and accepted a low offer without timely notice. The incident exposes risks from giving agents broad control over private data.
Cloudflare is building CryptoLabe to locate cryptographic code, trace dependencies and measure migration progress as it works toward full post-quantum readiness across its platform by 2029.
ElevenLabs’ new speech model follows cues more reliably, supports over 90 languages and keeps characters consistent across segmented productions. A Turbo version starts in about 150 milliseconds.
Nvidia’s Kumo Tabular predicts labels for new table rows without task-specific training or feature engineering. The commercially usable model comes in three sizes and supports classification and regression.