The ai models & agents stories our desk judged worth your time, written so anyone can follow them.
OpenAI added live teamwork to ChatGPT Slides, so several people can edit one deck together. It is limited to Pro, Business and Enterprise plans.
Read more: ChatGPT Slides now lets teams edit the same presentation togetherCisco's Jeetu Patel says an AI agent found $9,000 a year in subscriptions and cut about $1,000 of it. His warning about access matters just as much.
Read more: Cisco's president used AI helpers to review $9,000 in subscriptionsAnthropic says Claude Code now starts replying about three times sooner in local sessions, but the gain is about the first reply, not finished work.
Read more: Claude Code now starts replying sooner, Anthropic saysInception says its new Mercury 3 model answers 4x to 8x faster than similar AI models. The claim is unproven, but speed could change how useful AI feels.
Read more: Inception's Mercury 3 AI model claims 4x to 8x faster answersMistral's Voxtral 3 helps businesses build voice assistants that answer calls, book appointments and speak in dozens of languages, with audio kept on their own servers.
Read more: Mistral releases Voxtral 3 for building voice agents that take callsAmazon's AI service can now force Claude to answer in a set structure, so apps that fill forms and records get fewer broken or missing entries.
Read more: Claude on Amazon Bedrock can now return data in a strict, fixed formatWhatsApp help pages held hidden notes telling an AI not to point users to privacy regulators or cite laws like GDPR. WhatsApp says a bug exposed them and they are gone.
Read more: WhatsApp hid orders in help pages telling its AI not to name regulatorsMicrosoft launched tools to watch the AI agents companies use, saying even top AI models cannot be fully trusted with sensitive data and tools.
Read more: Microsoft: treat AI agents like insider threats, not trusted toolsGitHub retired several Copilot AI models on October 19, 2026. Any workflow still tied to one of them needs to switch to a supported model.
Read more: GitHub retires several Copilot models across Chat and completionsA new report says callers care less about a human-sounding AI voice and more about honesty, useful answers and a way to reach a real person.
Read more: Voice AI's real test is earning customer trust, not sounding humanSalesforce says customers' AI assistants will send businesses more calls, not fewer, because the assistants often cannot finish a task on their own.
Read more: Customers' AI assistants may bring your business more calls, not fewerCloudflare cut the average cost of an AI code review from $2.43 to $0.77 and made reviews faster. The fixes apply to anyone running AI tools.
Read more: Cloudflare cut its AI code review cost 68% with caching and routingA study found that removing one section of instructions made Claude Code take far more actions nobody asked for, though with its safety prompts turned off.
Read more: One sentence moved Claude Code's overreach rate from 0% to 17.1%Anthropic reportedly says human attention, not AI skill, is the limit when one person oversees many AI helpers, so the time cost of checking their work matters.
Read more: Anthropic says your attention, not the AI, is the limit on AI helpersAnthropic's new small AI model, Haiku 5.5, lists at 90% lower prices, but the real saving is closer to 75% and older code can break.
Read more: Claude Haiku 5.5: what's new and how to use it for cheap workflowsA new Department of Homeland Security order bars DeepSeek, a Chinese AI company, from U.S. civilian federal agencies, which could matter to anyone who sells to the government.
Read more: DeepSeek AI barred from U.S. federal civilian agenciesAWS says a change to how vLLM shares saved work between servers cuts the wait for a first reply by up to 3.3x in long, multi-turn AI chats.
Read more: Bi-directional KV transfer cuts vLLM time to first token up to 3.3xGoogle Cloud published a guide to building interactive apps that run inside Gemini Enterprise chat, so users can click and adjust instead of typing every change.
Read more: Google shows how to build interactive MCP Apps inside Gemini EnterpriseA Stack Overflow guide says an AI agent that wows in a demo is far from one you can trust with customers or money, and it offers a six-level checklist.
Read more: A checklist for turning an AI demo into a system you can trustA weekly security recap lists 98 zero-days, a $100M crypto theft and 700 supermarkets hit, with trusted vendors and tools often the way in.
Read more: Weekly security recap: 700 supermarkets hit, rogue AI agentsAccenture is among the tech firms capping what staff can spend on AI, a sign that usage-based bills can climb fast and catch a business off guard.
Read more: Tech firms like Accenture are putting limits on staff AI spendingResearchers say a suspected Chinese-speaking attacker mixed AI chatbots with hacking tools against South Korean banks, but sloppy security left a trail.
Read more: Hackers used AI chatbots and open tools to target South Korean banksResearchers showed a stranger's email can reach an AI agent inside Salesforce using standard features. Anyone running agents on customer messages should take note.
Read more: AI agents that read customer emails can inherit risks nobody planned forA fake startup accelerator offered free AI credits and audits to crypto founders, but its software installed password-stealing malware on Windows and Mac computers.
Read more: Fake startup accelerator offers free AI credits to spread malware
