1. Anthropic Published the Evidence, Then Called for a Pause — and Nobody Stopped Racing
Thursday was a genuinely surreal day in AI news. Anthropic published its Recursive Self-Improvement (RSI) report, showing that Claude now writes over 80% of its own code, produces 8× more code output than human contributors, and is closing in on the ability to improve its own core architecture without human oversight. Then on the same day, Anthropic’s Jack Clark called for a global pause on frontier AI training.
The same week, Anthropic also signed a joint letter to Congress with the CEOs of OpenAI, Google DeepMind, and Microsoft demanding synthetic DNA screening laws — because their own models could help build bioweapons.
Here’s the tension that doesn’t get enough attention: Anthropic has the data. It’s publishing it. It’s calling for regulation. And it raised $65B at a $965B valuation in the same window. The company that most urgently warns about AI risk is also scaling at maximum velocity. These aren’t contradictions — they’re the shape of an industry that sees the cliff and keeps accelerating.
Why it matters: The “AI pause” isn’t going to happen voluntarily. Every lab has the same data and the same incentives. The question is whether regulation arrives before the RSI timeline does. NZ’s AI regulation is still in consultation phase — which means we’ll be rule-takers on safety standards designed elsewhere.
The counterpoint: OpenAI’s Sam Altman reportedly dismissed the pause call as “marketing.” No major lab has signed on. If nobody stops, the pause is just a press release.
2. Bots Now Outnumber Humans Online — Agentic Traffic Grew 7,851%
Cloudflare and HUMAN Security both dropped data this week confirming what many suspected: bots now generate 53% of all internet traffic, with agentic traffic (AI agents operating independently) growing 7,851% year-over-year. Machine-to-machine traffic has officially surpassed human-to-machine traffic.
The HUMAN Security 2026 State of AI Traffic report found that automated traffic grew eight times faster than human traffic. AI agents are now transacting on the open web — booking tickets, filling forms, scraping content, and interacting with APIs — all without a human in the loop.
Why it matters: Every website that optimises for human visitors and ignores bot detection is already behind. CAPTCHAs are dead. Rate limiting is table stakes. The internet is becoming a machine-first infrastructure built by machines for machines — and that changes everything about security, ad economics, and content strategy.
3. Trump Signs AI Executive Order — Classified Benchmarks for Frontier Models
President Donald Trump signed an executive order on Tuesday creating a classified government benchmark for evaluating advanced AI models’ cybersecurity capabilities. Federal agencies have 60 days to develop the benchmark and set the threshold at which a system qualifies as a “covered frontier model.”
The order also establishes a voluntary framework where companies give the government access to models up to 30 days before release (down from 90 in an earlier draft). IBM CEO Arvind Krishna applauded the focus on AI security and open-source software.
The order arrives alongside the Pentagon labelling Anthropic a supply chain risk — a sign that the US government sees AI companies as both partners and potential threats.
Why it matters: The classified benchmark approach means the government is treating AI capability assessment as a national security matter — like nuclear weapons testing. But the 30-day voluntary window is a lot weaker than the 90-day mandatory review some wanted. The question is whether this becomes mandatory after the first incident.
4. Nvidia’s Jensen Huang Mounts Seoul Charm Offensive — Robotics, AI, and Samsung
Nvidia CEO Jensen Huang arrived in South Korea this weekend for a four-day trip that the local press is calling a “charm offensive.” He’s meeting with Samsung executives, appearing on Korean TV talk shows, and even attending a baseball game — all while positioning Nvidia’s role in Korea’s push into robotics and AI chip manufacturing.
The visit comes as Nvidia pushes deeper into robotics hardware (beyond GPUs) and Korea seeks to secure its position in the AI chip supply chain. Samsung’s memory business is critical to Nvidia’s supply, but the relationship has been tense since Samsung’s HBM memory qualification delays.
Why it matters: This isn’t just diplomacy — it’s supply chain management. Korea makes the memory that goes into Nvidia’s AI chips, and Huang needs Samsung’s HBM4 technology for next-gen Blackwell and Rubin architectures. A happy Samsung means stable supply.
5. OpenAI’s Sora Architect Now Leads a New Robotics Push — Building Physical Hardware
OpenAI has rebuilt its robotics division from scratch, with Aditya Ramesh — the architect behind Sora, the video generation model — now leading the effort. The company is hiring hardware engineers and signalling that it wants to build physical robots that work in the real world, not just software that manipises pixels.
Sam Altman has described a long-term vision of “personal robots for everyone,” and the team is starting with infrastructure tasks before moving to consumer products.
Why it matters: OpenAI tried robotics in 2021 and shut it down. This is a reboot with new leadership, more funding, and a clearer path. If Sora’s spatial reasoning can transfer to physical robot control, this could be OpenAI’s biggest pivot since Codex.
6. ChatGPT Memory System Rolls Out to US Users — “Dreaming V3”
OpenAI has started rolling out its updated ChatGPT memory system, internally called “Dreaming V3,” to US users. The new architecture provides fresher, more relevant personal context by intelligently prioritising which memories to keep and which to refresh.
The feature remembers user preferences, work history, and ongoing projects across sessions — making ChatGPT feel more like a long-term collaborator than a fresh start every time. Privacy controls let users review and delete specific memories.
Why it matters: Memory is the feature that turns ChatGPT from a chatbot into a personal assistant. This rollout signals OpenAI is serious about persistent context — but it also raises the stakes on privacy. NZ users will get it later, but the architecture choices set the pattern for everyone.
7. Florida v. Sam Altman — First-of-Its-Kind State Lawsuit Moves Forward
Florida Attorney General James Uthmeier’s historic lawsuit against OpenAI and Sam Altman personally is entering its early legal phases. The 83-page complaint alleges that ChatGPT has contributed to mass shootings, driven users to suicide, and engaged in deceptive trade practices — claims that would be extraordinary if proven.
OpenAI has not yet filed a formal response, and the case faces significant First Amendment and Section 230 hurdles. Legal observers expect a motion to dismiss in the coming weeks.
Why it matters: This is the first time a state has named an AI CEO personally in a consumer protection lawsuit. If it survives dismissal, it sets a precedent that could see every AI CEO facing personal liability for their product’s alleged harms — a nightmare scenario for the industry.
⚠️ Legal note: These are allegations in a complaint, not established facts. The lawsuit claims specific harms; none have been proven in court. OpenAI denies the characterisation.
8. NZ Data Centre Debate: Regional Benefits vs Resource Strain
Two stories this week capture the NZ AI data centre tension perfectly. Datagrid’s proponent Simon Currie told RNZ that AI data centres could drive regional economic development, attract investment in renewable energy and water infrastructure, and turn NZ’s clean electricity into a globally traded technology product.
But a separate report from 1News warns that data centres could produce as many emissions as the UK by 2030 and occupy land the size of Northern Ireland globally. The Spinoff also ran a deep dive on the real-world environmental cost.
Why it matters: NZ has clean electricity and a cool climate — genuine advantages for AI data centres. But every data centre pitch needs to answer the resource question honestly. Claims of “economic development” can’t paper over the power and water consumption reality.
9. Pentagon Labels Anthropic a Supply Chain Risk
The US Department of Defense has designated Anthropic as a supply chain risk under a Trump administration policy threatening to restrict ties between US defence agencies and certain AI companies. The designation, first reported by NPR, could limit Anthropic’s ability to contract with the Pentagon.
Why it matters: The irony is thick — Anthropic, the company most loudly calling for AI safety regulation, is deemed too risky for the Pentagon to buy from. It reflects the administration’s broader tension between embracing AI innovation and viewing AI companies as potential security liabilities.
U of Toronto Worm Paper — No Major CISA Response Yet
Last week’s bombshell — University of Toronto researchers building an autonomous AI worm on a free open-weight model — hasn’t yet triggered major government advisories. No CISA or NCSC alerts have been issued. No attribution of the specific model used.
Why it matters: The quiet response is almost as concerning as the research itself. If a worm that can spread through enterprise networks on a single GPU doesn’t trigger a government advisory, what would?
🔍 THE BOTTOM LINE
This was a week where three structural shifts converged: bots became the majority internet users, AI started building itself, and governments rushed to classify AI as a national security matter. Each story is big on its own. Together, they describe a world where the internet no longer primarily serves humans — and where the people building the most powerful systems are the ones calling loudest for brakes they refuse to install.