If last week was about money and plumbing, this week was about rules. Washington floated its most serious attempt yet at a national AI law, Anthropic admitted its own AI now writes most of its code and warned the technology may soon start improving itself, and Australia's signals intelligence agency quietly gained access to one of the most powerful — and most controversial — security tools yet built. Here's what mattered in the past seven days, and why it should be on your radar.

Congress unveils the "Great American AI Act"

On 4 June, Representatives Jay Obernolte (R-CA) and Lori Trahan (D-MA) released a discussion draft of the Great American AI Act, a sweeping 269-page bipartisan framework for how the United States would govern artificial intelligence. The headline — and most contested — provision is a three-year preemption of state laws covering how AI systems are built, while leaving states free to regulate how AI is used within their borders. The draft also adds penalties for using AI to impersonate government officials, requires large frontier developers to report critical safety incidents, directs the Census Bureau and Bureau of Labor Statistics to add AI-adoption questions to federal surveys, and authorises US$100 million a year for a Center for AI Standards and Innovation. As Roll Call reported, it is a discussion draft released to gather feedback before any formal bill is introduced.

Why it matters: The US has so far governed AI through a patchwork of state laws — Colorado's high-risk AI rules take effect this month — which is expensive and confusing for companies operating across state lines. A single federal standard would simplify compliance, but critics argue a temporary preemption could weaken protections by stopping states from policing how models are built. For Australian businesses, US rules tend to set the global tone: whichever way this lands, expect frontier-lab disclosure and incident-reporting to become standard expectations rather than nice-to-haves.

Anthropic warns its AI is starting to build itself

Also on 4 June, Anthropic published a report titled "When AI builds itself", disclosing that Claude authored more than 80 per cent of the code merged into Anthropic's own codebase in May 2026 — up from low single digits before its Claude Code tool launched in early 2025. As Tom's Hardware reported, the typical Anthropic engineer is now merging roughly eight times as much code per day as in 2024. The company's research arm cautioned that this points toward "recursive self-improvement" — the moment a model can design and build its successor with little human input — and argued the world should keep open a verifiable, multi-country option to slow or pause frontier development if needed.

Why it matters: This is a leading lab saying, in effect, that its own tools are accelerating its own progress, and that today's safety checks — designed for models that change between training runs, not during deployment — may not keep pace. The practical signal for business readers is twofold: AI coding tools are now delivering order-of-magnitude productivity gains in capable hands, and the people building these systems are openly uncertain about where the curve leads. That tension is exactly what to weigh when deciding how deeply to wire AI into critical workflows — a judgement call where having someone who can put models into production responsibly makes the difference between a useful tool and an unmanaged risk.

Australia's spy agency gains access to Claude Mythos

Anthropic expanded "Project Glasswing" — its initiative to find and fix critical software vulnerabilities — to around 150 organisations across more than 15 countries, including the Australian government. According to TechCrunch, the Australian Signals Directorate gained access to a preview of "Claude Mythos," an automated tool that scans software for serious bugs and security flaws. The program, which The Conversation analysed in detail, launched in April 2026 with founding partners including AWS, Apple, Google, Microsoft and NVIDIA, and its partners have reportedly surfaced more than 10,000 high- or critical-severity vulnerabilities.

Why it matters: This is the clearest sign yet that frontier AI is becoming national-security infrastructure — and Australia is now inside the tent. A system that can automatically uncover critical vulnerabilities is genuinely double-edged: a gift to defenders racing to patch their systems, and a danger if comparable capability reaches attackers. The Conversation's piece flags exactly this concern, noting the tool may improve cyber safety "but not for everyone." For local organisations, the takeaway is that AI-driven security testing is moving from research demo to deployed reality, and the bar for "secure enough" is about to rise.

Suno raises US$400m as the copyright fight rolls on

On 3 June, AI music startup Suno announced a US$400 million Series D at a US$5.4 billion valuation, led by Bond Capital — roughly doubling its valuation in about seven months. The company says it has passed two million paying subscribers, with users generating more than seven million songs a day. As Variety noted, the raise lands even as Suno fights copyright lawsuits from Universal, Sony and others over training on copyrighted recordings, though Warner Music settled and signed a licensing deal late last year.

Why it matters: Suno is a live test of the single biggest unresolved question in generative AI — whether training on copyrighted material counts as fair use. Investors are betting that it does, or that licensing deals will eventually bridge the gap. But the litigation could reshape the economics of every generative tool that learns from creative work, not just music. If your business uses generative AI for content of any kind, the sensible move is to watch where these legal lines settle and keep provenance and licensing front of mind.

Bezos backs a US$500m bet on brain-inspired AI

Flourish, a neuroscience-based AI startup co-founded by Internet Explorer creator Thomas Reardon, reportedly closed a US$500 million round at a US$2.5 billion valuation, with backers including Jeff Bezos, Lux Capital, GV (Alphabet's venture arm) and Catalio. Rather than scaling today's brute-force approach, Flourish is studying real neurons to find what it calls the brain's "core algorithm," aiming to build a system — "Cortex AI" — that runs on roughly 20 to 50 watts. That is laptop-level power, an order of magnitude below conventional AI hardware.

Why it matters: The dominant AI narrative is "bigger models, more compute, more power" — exactly the dynamic driving the data-centre and energy strain we examined in last week's roundup. Flourish is a bet on the opposite: radically more efficient AI inspired by biology. It is early and entirely unproven, but if energy turns out to be the real ceiling on AI, then efficiency breakthroughs could matter as much as raw capability — and reshape which AI is actually affordable to run at scale.

Quick hits

  • OpenAI builds out Codex for business. OpenAI added enterprise-focused features to Codex, including business plugins spanning sales, data analytics, creative production, product design and finance, plus document annotations that let users target and revise specific sections (OpenAI release notes).
  • Alibaba's Qwen keeps the pressure on. Alibaba's Qwen 3.7 line continues to push open-API models toward frontier performance — a one-million-token context window and strong agentic benchmarks at a fraction of closed-model pricing (OpenRouter). If you're weighing which models to test, our rundown of AI tools and resources is a useful starting point.
  • Generalist AI raises US$400m for robotics. The startup picked up US$400 million at a reported US$2 billion valuation to build AI that lets robots handle complex physical tasks (Crunchbase).
  • Meta tests AI for your lapel. Meta is reportedly developing an AI-powered pendant and a business offering called "Wearables for Work," alongside an expanding line of AI glasses (industry reports).

Looking ahead

Apple's WWDC keynote kicks off today, 8 June, with a long-awaited Siri overhaul widely expected to lean on Google's Gemini — a notable admission from a company that prefers to build in-house. Watch, too, for Google's Gemini 3.5 Pro, signalled for release this month, and for SpaceX's IPO, reportedly priced for 11 June with trading to follow — an early market test of investor appetite for the AI-adjacent mega-listings now lining up.

That's the week in AI. The headlines were about lawmakers, labs and large cheques, but the practical story hasn't changed: this technology only pays off when it's matched to a real problem and deployed with care. We'll be back next week with the developments that matter — see you then.