AI2Day Daily Brief — 10 Aug 2026

Good morning — it is Monday, 10 August, and this week opens with governments, hedge funds, and tech giants all wrestling with the same basic question: who or what should actually be in charge when AI is doing the work.

We start in New South Wales, where Deputy Premier Prue Car has written to the NSW Education Standards Authority — known as Nesa — asking it to urgently examine how artificial intelligence is affecting student learning and to consider banning unsupervised take-home assessments. The stakes are real. Half of every student's Higher School Certificate result comes from school-based work, and some of that is completed at home with no supervision. Car's concern is direct: if a student can feed a question to an AI chatbot and paste the answer into an assignment, the mark no longer reflects what they actually know. She also cited growing evidence that AI use is affecting students' ability to think and reason independently. No ban is in force yet — Nesa must finish its review before any rules change.

A timely tension: the same AI that worries educators is, according to Anthropic, better at catching its own mistakes than humans are. Starting 14 August, Claude Code — Anthropic's tool that writes, edits, and runs code on your behalf — will switch to auto mode by default for users on Pro, Max, and Team plans. Auto mode means the tool works on its own and only stops when an action is classed as irreversible, destructive, or aimed outside your own systems. The reason Anthropic gives for removing routine human approval is a study from their own testing: human reviewers approved 97 percent of all permission prompts they saw, catching only 13.6 percent of harmful actions. Auto mode, by contrast, caught 89 percent. Anthropic is also adding prompt injection screening and customisable hard-deny rules alongside the change.

Worth sitting with that finding before you click approve on anything today.

Are you at risk? If you own or run a business, one wrong click is all it takes. Train2Secure trains your employees to spot phishing emails and scams before they fall for them, with short lessons they will actually finish. It starts from just $1.59 per user, per month, way less than a cup of coffee. Head to Train2Secure dot com for a free trial today. That's Train, the number two, Secure, dot com.

Moving to markets. The hedge fund Situational Awareness — founded in 2024 by Leopold Aschenbrenner, a researcher previously at OpenAI — has invested 400 million dollars in chip startup Source Foundry this week, bringing its total stake in the company to 500 million dollars. Source Foundry was founded by Stanford researchers and is working to make semiconductor chip manufacturing faster and cheaper. The investment is notable given the fund's recent turbulence. Situational Awareness saw its assets under management fall from 20 billion to 10 billion dollars after steep losses in AI infrastructure stocks, and at the end of July it sold most of its public stock holdings to Ken Griffin's Citadel — though it held on to its shares in Anthropic. A fund that just halved in size writing a 500 million dollar cheque into chip manufacturing is a signal worth watching.

Also worth watching: where the US and China actually stand in AI. This week, Hugging Face chief executive Clément Delangue told CNBC that China is, in his words, clearly dominating on open models right now, and that he would not be surprised if Chinese labs reach the very frontier of AI capability by the end of this year or in 2026. That is a strong claim, so what does the evidence actually show? China leads clearly in two areas: open-source models, where every major freely downloadable model currently comes from a Chinese lab, and robotics. Chinese models also compete aggressively on price. Where the US retains advantages is at the top tier of model performance, in private investment capital, and in specialist talent. US export controls continue to limit China's access to the most advanced chips, which slows training of the largest models. Experts also note that if Chinese AI becomes the default technology for developing nations, that carries political weight well beyond any benchmark.

And finally, Tencent has open-sourced a shared memory system for teams of AI agents, called Team Memory, now in beta. The idea is straightforward: instead of each AI agent keeping its own separate notes about a user or a project, every agent on a team reads from one central pool. On Tencent's own accuracy benchmark, adding a persistent shared profile lifted correct responses from 48 percent to 76 percent — a significant jump. The GitHub repository hit number one on the TypeScript trending list within days of launch. The problem Tencent has not yet solved is equally straightforward: if a wrong fact enters that shared pool, it reaches every agent at once, and Tencent's own documentation includes no process for correcting or expiring a memory item that turns out to be wrong. Useful architecture, unresolved failure mode.

For a deeper look at all five stories and the broader week in AI, join us for AI Today Weekly — our longer Monday show. That is your AI briefing for today. Every story is at A-I-2-Day dot live. That is A, I, the number two, D-A-Y, dot live. We are back tomorrow morning. If you got something out of this, a thumbs up and a subscribe genuinely helps.