AI Titus's Morning Wire

AI NEWS TODAY

HACKERS ARE ACTIVELY STEALING CLOUD LOGINS THROUGH A CRITICAL HOLE IN MLFLOW, A TOOL THOUSANDS OF AI TEAMS RUN. PATCH NOW, SAYS CISA
★ Must-Read / Watch agentic engineering / build-better-agents📚 Learn state of AI coding / useful technique

⚡ AI Frontier · smol.ai

★ Must-ReadChatGPT on Mac can now read, search and send your text messages
Announced August 20 for ChatGPT Work and Codex plans. It connects to Apple Messages, so it can summarize a group chat, draft a reply, and send iMessage, SMS or RCS for you. The catch is the permission grant: full disk access plus your contacts plus the power to send messages as you. Genuinely useful, and exactly the kind of access worth pausing on before a whole office enables it.
★ Must-ReadCursor's cloud coding agents now watch PRs, Slack and cron jobs on their own
The August 19 changelog lets a Cursor agent subscribe to a pull request, a Slack thread or a schedule and keep working until the goal is met, including a new open-ended goal command. The direction is clear: coding agents are moving from a chat window you sit in to background workers that act while you are elsewhere. Worth trying on one repo before trusting it with many.
📚 LearnLinear's AI agent can now run your code in a real sandbox and test it in a live browser
Shipped August 20. Instead of handing back untested code, the agent spins up a working environment, runs the change, and clicks through it in a browser before you see it. Verification built into the loop is the difference between a demo and a coworker, and this is one of the first mainstream tools to build it in.
Grok 4.6 is now generally available on Amazon Bedrock with a 500K context window
From August 19, companies already on AWS can use xAI's flagship model through Bedrock without a separate vendor relationship, with a 500,000 token window, roughly a small bookshelf of text per request. The practical point is procurement: every major model is becoming available inside every major cloud, so switching costs keep falling.
📚 LearnThe first independent report card on AI lab safety controls: nobody scores above a C+
GuideLight, a nonprofit not funded by the labs, graded Anthropic, OpenAI, Google, Meta and xAI on whether they could actually detect and stop their own systems misbehaving. Top grade: C+. Read it before repeating any vendor's safety marketing, and note the gap between what labs promise and what an outside grader can verify.
★ Must-ReadAmazon will bring 30-minute drone delivery to nearly 500 US cities by year-end
Announced August 19: a sixfold jump from around 11 metro areas today, free on Prime orders over $50. These are AI-piloted aircraft flying themselves over neighborhoods at scale. Autonomy stops being a demo the moment it lands in your yard with the toothpaste.

📺 Watch · latest videos

📚 LearnWF2026: Autoresearch & Keynotes ft. Anthropic, Google DeepMind, Amazon AGI, Sonar, Arena, Recursive
Featured by Latent Space.
AI Engineer
📚 LearnThe State of Model Routing — NVIDIA, Cognition, OpenRouter
Featured by Latent Space.
AI Engineer
📚 LearnVercel accuses Cloudflare of stealing
Featured by Latent Space.
Theo - t3․gg
Turn Claude Into a One Person Marketing Team in 38 Mins
Latest from Nate Herk.
Nate Herk · 812 views · 64 likes
The Missing Layer: Design Taste in AI Agents — Hassan El Mghari, Together AI
Most people can spot a vibe coded app in two seconds and cannot say why. Hassan El Mghari names it: the purple gradient background, italics in the hea
AI Engineer · 1 views · 3 likes
Grok Bot can shop for you! (I hope it doesn't though)
Latest from Matthew Berman.
Matthew Berman · 7K views · 57 likes
DeepSeek is back... and Silicon Valley is terrified
Take BlueDot's Future of AI course for free - https://bluedot.org/fireship OpenAI just paused its biggest training run in history just as things were
Fireship · 566.1K views · 20.6K likes
DeepSeek Just Built the Next Generation of Coding Agents
Latest from Cole Medin.
Cole Medin · 17.1K views · 532 likes
★ Must-WatchHow to choose the right Claude model for any task
Choose a Claude model based on the complexity of the task. Then adjust the effort level to move between faster or more thorough results.
Claude / Anthropic · 6.7K views · 304 likes
★ Must-WatchHow to Turn a Business Question Into a Strategy Deck With ChatGPT Work | Tutorial
Turn a business question into a clear recommendation—without spending days gathering research and building slides. In this demo, Arvind from OpenAI’s
OpenAI · 7.3K views · 194 likes

💻 Builder Desk · Hacker News

HTML Can Do That
888 pts · 201 comments
I like 'em thick: an apology to my English teachers
781 pts · 320 comments
The August 17 outage
577 pts · 637 comments
Malicious Rust crate Arrayref runs a build-time payload
519 pts · 448 comments
CIA funding helped keep NeXT afloat in the 80s
424 pts · 257 comments
I should have loved biology (2020)
304 pts · 115 comments

🗣 Voices & Blogs

★ Must-ReadZvi Mowshowitz's AI #182: Pause For Reflection
The weekly close read, August 20. His frame this week: the industry just had a rare quiet stretch to digest the recent shocks, and he sorts what actually changed from what was noise. If you read one roundup to stay current without living on X, this is the one.
📚 LearnSimon Willison: ChatGPT search quietly started using site-specific queries at scale
August 20. He measured ChatGPT's search jumping from site-targeted queries 0.4 percent of the time to about 17 percent almost overnight. If customers finding your business increasingly means an AI finding it first, this is the kind of shift that changes your traffic without warning.
GuideLight's control assessment, the primary document behind the C+ grades
The full graded assessment of five frontier labs by an independent nonprofit. Worth skimming the methodology, because how you would even check an AI company's safety claims turns out to be the hard and interesting part.
Dream's situation report: state-linked hacker groups are using autonomous AI agents
A practitioner recap of research showing nation-state groups running agentic AI in live operations, still being discussed this week. Pairs with today's siren: AI is now on both sides of the security equation, doing the attacking as well as being the target.
Most coding agent benchmarks skip large-scale refactoring. Not this one.
AI coding agents still struggle with large-scale refactoring, with the best model achieving only a 41.2% resolve rate on a The post Most coding agent… · The New Stack

📦 What's Being Built · GitHub

Significant-Gravitas/AutoGPT
AutoGPT is the vision of accessible AI for everyone, to use and to build on. Our mission is to provide the too · ★186.7K · Python
ollama/ollama
Get up and running with Kimi-K2.6, GLM-5.2, MiniMax, DeepSeek, gpt-oss, Qwen, Gemma and other models. · ★179.1K · Go

🌏 The Wire · Drudge / Breitbart

ANTHROPIC MOVES COMPUTER USE, SKILLS API AND FILES API TO GENERAL AVAILABILITY FOR PRODUCTION AGENTS
August 20. Claude agents that click and type inside real apps and browsers are out of beta, alongside the Skills and Files APIs. The building blocks for do-it-for-me software are now sold as production plumbing, not previews.
GOOGLE LOCKS IN AN OPTION TO BUY UP TO $12.2 BILLION OF MARVELL STOCK IN AN EXPANDED AI CHIP DEAL
Straight from the August 19 SEC filing. Google is tying itself to a second custom-silicon partner rather than betting everything on one supplier. The AI chip market keeps consolidating around whoever the giants pre-commit to.
SALESFORCE OPENS ITS AGENT PLATFORM WIDER WITH NEW MCP SERVERS AND SLACK HOOKS
August 19. Agentforce agents get standard MCP connections into more business systems plus deeper Slack integration. The enterprise agent race is now about who can act across the most systems safely, not who chats best.
PERPLEXITY'S AI WILL RUN TASKS STRAIGHT FROM YOUR INBOX. JUST CC IT ON AN EMAIL
CC computer@perplexity.com and its agent picks up the task from the thread. Email is the oldest workflow there is, and the assistants are coming to it rather than making you leave it.
AI NOW WRITES JUST UNDER HALF OF ALL ISSUES CREATED IN LINEAR, AND THE SHARE IS STILL CLIMBING
Linear's own live data, August 2026. Issue tickets written by agents are approaching parity with ones written by people. Quiet numbers like this say more about real adoption than any launch keynote.
OPENAI PARTNERS WITH SECURITY FIRM ABNORMAL AI TO WATCH FOR AI AGENTS GOING OFF SCRIPT
August 19. Abnormal, which protects over 4,500 organizations, joins OpenAI's cyber partner program to build tools that monitor agent behavior and flag deviations. A whole vendor category is forming around the question: what is your AI actually doing?
NVIDIA'S SEC FILING PUTS ITS GUARANTEE FOR OPENAI'S OHIO MEGA DATA CENTER AT $105 BILLION, FAR BELOW THE RUMORS
The filed number is $145 billion less than the $250 billion that circulated in July. Still a staggering commitment to a single campus, and the circular-financing question, chip maker guaranteeing its own biggest customer, is not going away.

🤖 Trending Models · Hugging Face

MiniMaxAI/MiniMax-Music3
text-to-audio · ★1.1K · 15.7K dl

📈 Markets

NVDA 217.56 ▼1.0%
MSFT 484.31 ▲0.6%
GOOGL 344.72 ▲0.2%
AMZN 265.84 ▲2.5%
META 546.03 ▲0.4%
AMD 466.42 ▼3.7%
AVGO 362.48 ▼4.6%
PLTR 175.19 ▲2.1%
SPCX 139.65 ▼2.6%
TSLA 351.12 ▲4.2%

The BriefMLflow is a free tool many companies use to track and manage their AI models. On August 19 the US cyber agency CISA confirmed that attackers are actively using a critical flaw in it (CVE-2026-64849) to reach into servers and pull out cloud passwords for AWS, Azure and Google, with no login needed. Any MLflow older than version 3.15.0 that can be reached from the internet is exposed right now. This matters to ordinary businesses, not just AI labs: AI tooling is becoming normal infrastructure, and it needs the same patching discipline as everything else on the network.

Level UpBookmark CISA's Known Exploited Vulnerabilities catalog and make checking it part of your patch routine this week. It is a free, searchable list of the security holes attackers are actually using in the wild, not just the theoretical ones. Search it for every AI or ML tool you or your clients self-host, MLflow included. Ten minutes against that list tells you what truly cannot wait. The CISA Known Exploited Vulnerabilities catalog, the patch-first list to check