Curiosities
DoorDash received only hundreds of orders placed by AI agents
Volume from agents like Muse, Instinct and Fo contrasts with an average of 3 million deliveries per day.
Search
15 result(s) for “AI agents”
Tools
DeepSeek proposes AI super-agent to coordinate multiple coding agents
System uses a central agent to split tasks and reduce human supervision in race against Claude Code and Codex.
Security
OpenAI agents talked inside the chip and broke isolation with no network
In a controlled test, agents in separate microVMs swapped fake secrets by timing CPU cache access.
Security
AI agents raised their own crypto spending limit in test
Nine out of ten AI models raised their own cap in Circle wallets by fetching approval codes from email.
Big tech
OpenAI trains its most advanced models without internet after agent incidents
Company built an offline environment with cached content and will only restore network access when new safeguards justify it.
Big tech
Nvidia invests US$ 3 billion in Meta to supercharge AI with Vera Rubin chips
Stock investment backs plan to raise up to US$ 95 billion for data centers and boost Muse models and autonomous agents.
Benchmarks
Google launches PLE, benchmark with 85 real engineering tasks for AI agents
Test with 30 mechanical, 27 electrical, 18 civil and 10 structural tasks measures deliverables, and the best AI scored 48.3% on the hidden set.
AI models
Google releases Gemini 4 Argon for coding, finance, law and science
Frontier model is now available globally via API, Gemini app and Search, with introductory pricing of US$ 2 for input and US$ 10 for output per million tokens.
Tools
Anthropic announces Claude Duet, AI agent for work tasks
The new tool runs on a virtual machine, asks for approval before acting, and rolls out first to Pro and Max subscribers on a waitlist.
Tools
Opera creates WebAI standard to run AI directly in the browser, without the cloud
Preview feature in Opera One uses open-weight models like Google's Gemma and processes everything on-device for greater privacy.
AI models
Microsoft launches MAI family with 7 AI models for reasoning, voice, code and image
The highlight is MAI-Thinking-1, the company's first reasoning model, with 35 billion active parameters and a 128K window.
Tools
Anthropic launches Claude Duet, agent that works alone and pings you when it needs you
Feature is in research preview in the United States for Pro and Max subscribers via waitlist and targets professionals and small businesses.
Security
CZ warns of more AI attacks after US$ 6.3 billion theft from Binance
FBI blames North Korean hackers for the theft, but investigators found no AI in the laundering of the funds.
Benchmarks
Intelligence Index v5 puts Claude Opus 5.5 on top and widens open-weight gap
New methodology uses 60% private tests and leaves Astra at 57 and Sol at 55 points behind the leader, according to Artificial Analysis.
AI models
42AI emerges from stealth with $62.5M to predict human behavior
Israeli startup builds a "behavioral world model" already tested at brokerages to anticipate traders' decisions.