Daily Briefing
22 August 2026 · 5 stories
🔥 Top story
1Anthropic's "virus detector" is not for you: it scans your code on your behalf
Picture a doctor who examines your computer, spots the flaw, and writes you the prescription — without ever letting you touch the stethoscope. That is what Claude Security, Anthropic's next-generation antivirus, now does. Under the hood runs Claude Mythos 5, the most powerful model the company has ever built for hunting software bugs. But you will never get to talk to it: Mythos is not for sale, not for rent, not accessible. It works backstage, inside Claude Code, and only hands you a report: "here is the flaw, here is the risk, here is how to fix it." To keep face with the open-source community, Anthropic is also adding $35 million of API credits to a fund for free-software maintainers. If you use a popular framework, your next patch may quietly be paid for by Anthropic without you knowing.
2Alibaba taught an AI to click on your phone for you — and it beats the American ones
Dreaming of an assistant that books your flights, fills your forms, moves your files while you sleep? Alibaba built exactly that: Qwen-UI-Agent, a "foundation" model that drives a phone, a computer, or a browser like a human. On a standardized exam against GPT-5 and Claude Opus, it scores 82% — ten points ahead of its American rivals. Trained on more than a hundred physical devices and a hundred and fifty long trial-and-error sessions, it also knows when to stop and ask for confirmation before sending an email or paying a bill. It is a quiet but massive shift: while the US debates agent safety, China ships a product that already knows how to use your mouse.
3DeepSeek gives vision back to its tiny models — for the price of one coffee per million words
Read a photo, understand a diagram, analyze a screenshot: what big models have done for two years, DeepSeek's "Flash" family just learned — for almost nothing. The new V4-Flash-Vision-Exp accepts images and text together, and charges 22 cents of a dollar per million input words — less than an espresso. Better still: no surcharge for images, while competitors bill each processed photo separately. Benchmarks place it on par with Anthropic's Opus 4.8, the most expensive model on the market. For a small business that wants to automate invoice sorting or receipt reading, it is time to look at China again.
4Anthropic releases its "recipe format" for AI agents — Microsoft and Google are already on board
If you have ever given a cooking recipe to a friend, you already know how to explain step by step what needs to be done. That is exactly the format Anthropic just standardized for AI agents: a small text file called a "Skill" that describes, in plain language, how the agent should behave for a given task. Since August 20, this toolkit is officially available on the Claude platform, and Microsoft has already added it to Foundry; Google Cloud has announced it for soon. In practice: if you write a "summarize a contract" Skill, it will run tomorrow both on Microsoft and on Google, without rewriting. For software robots, this is the equivalent of the PDF format arriving: everyone aligns on the same convention.
5Devin, the agent that codes for you, could soon be worth $40 billion — but the money is not signed yet
Cognition, the company behind Devin (the agent that writes computer code like a junior engineer), is talking to investors about raising more than a billion dollars at a $40 billion valuation — double the figure from May. In just a few months, its revenue has reportedly nearly doubled, approaching a billion a year. Its customers? Mercedes, NASA, Goldman Sachs, Cisco. If the round closes, it would be one of the highest valuations ever granted to an AI agent dedicated to code, on a par with Cursor or with Microsoft Copilot. For now, the amount and the valuation remain market rumors: nothing is signed.
📡 To watch
Hermes in the Dream report — Info-Sys must check it is not ours
A Taiwanese cybersecurity report from August 12 lists "Hermes" among the frameworks used by attackers. OpenClaw, the other name cited, has already been verified: it is a public project unrelated to Info-Sys. "Hermes" still needs to be checked urgently: confirm it is a namesake, and not a misattribution to our hermes-agent.org project.
Skills: 15 channels in parallel to distribute one agent
Anthropic released its "Skills" format in final version on August 20. Microsoft Foundry already distributes it; Google Cloud's Vertex AI has announced it for soon. For an agent publisher like Info-Sys, a single Skill file becomes simultaneously distributable across at least fifteen partners — the equivalent of putting a product on fifteen store shelves in one single write.
DeepSeek V4-Flash-Vision-Exp → stable version in the coming weeks
The "exp" suffix means experimental version. Historically, DeepSeek promotes its experimental models to stable within two to four weeks. If the promotion lands, today's price — 22 cents per million words — becomes the reference price for low-cost multimodal.
Cognition / Devin: the $40 billion round still has to close
The funding round reported by Bloomberg on August 13 is in negotiation. Amount and valuation can change at closing. If the round signs at $40 billion or more, it would be the largest investment ever made for a code-dedicated agent — a signal that agent-market valuations are catching up with those of model labs.
📊 Trend
This week, two forces are pulling against each other. On one side, the big labs lock their most powerful models behind scanned products (Anthropic Mythos) or youth programs (ChatGPT for Teens), refusing to expose them directly. On the other, Chinese publishers ship multimodal models and full agents in days, at prices ten times lower. For an everyday user, the consequence is simple: agents are becoming affordable and multilingual, but the cutting-edge models stay out of reach — unless you go through a large company that negotiates access for you.