Daily Briefing
17 August 2026 · 5 stories
🔥 Top story
1Microsoft releases the invisible layer that powers AI agents
Imagine a virtual assistant that doesn't just answer questions but also books your flight, checks your calendar, and places an online order while you sleep. To do all that reliably, it needs an invisible technical layer: a harness that handles memory, sequences steps, and secures access. Microsoft has just released this harness as a stable product, in the same vein as what Amazon, Google, LangChain, and NVIDIA already offer. For small businesses that want to automate tasks without rebuilding their IT, it's now a standard brick to plug in.
Source: https://devblogs.microsoft.com/agent-framework/microsoft-agent-framework-at-build-2026-announce/
2DeepSeek open-sources the engine that drives AI agents
DeepSeek, the Chinese lab that surprised everyone with its cost-effective model back in late 2024, is taking a new step: it is freely releasing a framework to orchestrate AI agents, under the MIT licence (the most permissive that exists). The slogan is clear: everything is a plugin. In practice, a company can assemble its own automated assistant from interchangeable building blocks, compatible with models from OpenAI, Anthropic, Google, or DeepSeek itself. It is the free, open-source equivalent of what Anthropic charges for with Claude Code. For technical teams, it is one more option to avoid depending on a single vendor.
Source: https://www.deepseek.com/harness/en/
3OpenAI pauses its most powerful model: it knows how to attack too well
Imagine software so good at spotting security flaws that it becomes a weapon in the wrong hands. That is what OpenAI discovered when testing its next model, Astra. For the first time, an AI model has reached the "critical" level on OpenAI's internal scale: it can find and exploit security flaws without any human help. Development has been frozen. OpenAI says it will take the time to put serious safeguards in place before any release. For the general public, the message is simple: the most serious labs are starting to voluntarily slow down when their creations become dangerous, and that is good news for everyone's security.
Source: https://openai.com/index/responding-next-frontier-critical-cyber-capabilities/
4A Chinese AI model finds a serious flaw in Cursor on the day of its release
Cursor, the AI-powered code editor used by millions of developers, has a serious flaw: an attacker can take control of a computer simply by feeding a text instruction to the AI. The vulnerability, nicknamed "CurXecute", was found on the very day a new Chinese model specialised in cybersecurity, GLM-5.3 from lab Z.ai, was released. The model was designed to spot flaws, and it found a big one in Cursor. The lab has decided to delay the free release of its model to give everyone time to verify safety. For Cursor users, it is another reason not to trust your AI assistant blindly.
Source: https://z.ai/blog/glm-5.3
5Alibaba releases a free compact AI model that punches above its weight
Imagine a city car with the performance of a luxury sedan. That is what Qwen3.8-27B does, a 27-billion-parameter model that Alibaba has just released for free under the Apache 2.0 licence. It reads text and images equally well, keeps track of a conversation across 262,000 words at once, and passes 84% of a test where it has to operate a computer like a human. It runs on a single high-end GPU. For a small business or freelancer who wants a powerful AI agent without sending their data to a cloud giant, this is a usable option starting today.
Source: https://huggingface.co/Qwen/Qwen3.8-27B
📡 To watch
Databricks raises $5 billion for its AI agents
The data platform Databricks closed a record $5 billion round on 13 August at a $190 billion valuation. Management confirmed it wants to bet on "enterprise agents", the AI assistants that work at the heart of large companies. For the industry, it is a sign that capital is concentrating on the platforms that build and deploy agents.
Anthropic makes the invisible traces of Claude visible
Anthropic detailed on 14 August how the digital watermark it embeds in texts produced by Claude works. It is a discreet marker that lets you recognise AI-generated text. The move is part of a wider industry effort to trace content before regulators require it.
Qwen3.8-Max still not available
Announced for the week of 10 August, Alibaba's large Qwen3.8-Max model is still not downloadable as of 17 August. That is J+24 from the initial promise. In the meantime, the compact 27B version has been delivered. To monitor daily on the lab's Hugging Face page.
GLM-5.3: free release delayed for security reasons
Z.ai's GLM-5.3 model, announced on 14 August, will not be released as open source for two weeks to give time to verify it cannot be used as a weapon. It is an unprecedented case for a Chinese lab. To follow: the actual release date and the licence chosen.
📊 Trend
The invisible layer of AI agents is taking shape at full speed. While the most powerful models are being paused out of caution, Chinese and American labs are working on the same missing piece: the "engine" that makes agents run reliably and securely. Six major editors now offer a version of it, three of them as open source. For companies, this means they now have a choice between a turnkey service and a free toolkit to assemble themselves. The general public will mainly remember that publishers are starting to voluntarily slow down when their creations become dangerous: it is a rare sign of responsibility in tech.