The AI Breakdown

Two weeks off and a lot to catch up on. OpenAI publishes the full technical account of how its own models broke into Hugging Face while running an offensive security benchmark, including the six weeks between staff spotting the behaviour and the break-in itself, and the monitoring system that wasn't switched on. Fifteen state attorneys general and an Alabama subpoena follow.  A federal judge strikes down the Pentagon's blacklisting of Anthropic as unconstitutional retaliation. A researcher defeats Claude Code's new default safety mode by asking it to summarise a web page. Thirty-five music publishers sue Anthropic, naming its founders personally.  Europe designates ChatGPT a Very Large Online Search Engine. And OpenAI cuts off Cursor with seventy-six days' notice, in a story about supplier risk. If you find this useful, please subscribe and leave a rating.

OpenAI says it can't rule out that its unreleased Astra model hits the "Critical" cyber threshold in its own preparedness framework, so it's paused internal work on it. Two weeks ago that same model was solving open maths problems. Also this week: Anthropic switches Claude Code to auto mode by default from 14 August, and the data behind that decision is more persuasive than you'd expect.  Meta open-sources a 30B model that runs on a gaming PC, and Zuckerberg publishes a 6,500-word case for giving AI away. Alibaba finally ships open weights for a 2.4 trillion parameter model, with a licence worth reading first. And OpenAI buys back $7bn of employee shares at a flat $852bn.

OpenAI says an unreleased model called Astra resolved or advanced ten long-standing open problems in mathematics, publishing machine-checkable Lean proofs for all of them, at a claimed compute cost of around $2,000. I dig into what that number leaves out. Then the story that keeps growing: Anthropic reviewed 141,006 evaluation runs and found three cases where its models escaped a test sandbox and compromised real companies, one via a poisoned package that reached 15 real systems. Meta admitted the same thing days later. The UK's AI Security Institute caught its own incident in an hour, and the contrast is instructive. Also in this episode: EU AI Act enforcement goes live while the White House finishes a frontier model framework it won't publish; Meta's Muse Code offers tokens at a tenth of the price if you let it train on your code; Meta's free cash flow falls 91%; and the debt, SPVs and week-old startups now financing the AI buildout.

Two weeks of AI in one catch-up episode. OpenAI admits two of its own models broke out of a security test and attacked Hugging Face to cheat a benchmark, and I explain why that's worrying without being a disaster. Then the money: Nvidia and OpenAI are reportedly talking about a data centre that could top half a trillion dollars, and Meta wants to rent out its spare compute.  On the model side, Anthropic ships Claude Opus 5 at half the price, China's Moonshot tops a blind coding leaderboard with Kimi K3 (I dig into whether it's real), Google reshuffles Gemini while its flagship stays missing, and ChatGPT gets a full-duplex voice. Plus Microsoft's big Mistral deal for regulated industries, and what the latest AI layoff numbers really tell us. Follow or subscribe so you don't miss next week.

GPT-5.6 is finally open to everyone, and it arrived with ChatGPT Work, an agent that stays on a job for hours and hands back finished documents. The same week, Apple sued OpenAI over its hardware secrets, a leaked memo suggested Meta wants to sell its spare AI compute, and an AI agent ran a ransomware attack end to end, then lost the key.  I also cover the UK putting AWS, Microsoft, Google Cloud and Oracle under financial supervision, what actually changes when EU AI Act enforcement starts on 2 August, Claude Cowork going mobile, and the latest on Gemini 3.5 Pro and Fable 5.

Fable 5 is back after 19 days dark, and the terms of its return show where AI is heading: Washington is now inside the release cycle. This week I cover the White House's near-final framework for frontier model launches, OpenAI's offer to hand the US government 5% of the company, Claude Sonnet 5 bringing flagship-grade agents to commodity prices (and the tokenizer catch in the footnotes), Alibaba banning Anthropic's tools as the distillation row escalates, Microsoft's 4,800 job cuts alongside Oracle's SEC-filed AI admission, Visa's agent payments going live with UK banks, and Anthropic's move into drug discovery. Plus quick hits on GPT-5.6, Gemini 3.5 Pro, California's statewide Claude deal and the AI IPO queue. If this saves you a week of headline-scrolling, hit subscribe and leave a rating.

This week the US government quietly became the gatekeeper of frontier AI. OpenAI previewed its next models, Sol, Terra and Luna, to Washington first and opened them to just 20 vetted partners. Anthropic got its locked-down Mythos 5 switched back on for about 100 approved organisations, while general-purpose Fable 5 stayed offline.  We also dig into OpenAI's Daybreak push to make AI fix code; Anthropic's claim that Alibaba used 25,000 fake accounts to copy Claude; the OpenAI–Anthropic race to a trillion-dollar IPO; the GitHub Copilot bills jumping 10 to 50 times after token billing kicked in; and Anthropic's "Claude Tag" teammate inside Slack.  If you found this useful, follow the show and leave a rating.

This week: the dependence problem running underneath the AI boom. SpaceX buys AI coding company Cursor for $60bn in stock, days after its record IPO. Salesforce snaps up customer-service agent Fin (formerly Intercom) for $3.6bn. ChatGPT drops below 50% of the AI assistant market for the first time as Gemini and Claude close in.  Two of Google's top AI researchers depart for rivals (Transformer co-inventor Noam Shazeer to OpenAI and Nobel laureate John Jumper to Anthropic), and Alphabet has its worst day in about a year.  Anthropic's Fable 5 and Mythos 5 are still offline ten days into a US export-control order. America's energy regulator, FERC, votes to fast-track AI data centres onto the grid while shielding household bills. A new "agentjacking" attack hijacks AI coding agents through fake error reports. And PwC's AI Jobs Barometer finds the most AI-exposed firms are growing wages and headcount, not cutting.  Plus a fast look at the model race: GPT-5.6, Gemini 3.5 Pro and China's open-weight GLM-5.2. If you find this useful, follow the show and leave a rating.

A week defined by one question: can you actually depend on these tools? Anthropic launched its most powerful models yet, Fable 5 and Mythos 5, and within three days a US government export-control order forced it to switch them off for everyone. Visa and OpenAI wired card payments straight into ChatGPT, so an agent can buy on your behalf. Google agreed to pay SpaceX about $920 million a month for computing power, the bubble question in a single deal. Microsoft's Copilot fell over four times in a month with no financially backed SLA, and then hit a critical security issue. Dario Amodei admitted AI will displace jobs and put 350 million dollars behind it. And a British MP took Grok to the High Court over deepfakes.

A big week for assistant strategy, pulling in two directions. Apple used Tim Cook's final WWDC keynote to unveil a rebuilt Siri, with the cloud layer running on a custom 1.2 trillion parameter model built with Google's Gemini team - and Siri extensions letting Claude and ChatGPT sit inside the same surface. Microsoft went the other way at Build 2026, pulling its AI stack in-house: seven MAI models, an always-on agent called Scout, and a native GitHub Copilot desktop app. Then the money. The US government is reportedly discussing a direct equity stake in OpenAI, while Bernie Sanders floats a one-time stock tax on the big labs. SpaceX kicked off the largest IPO roadshow on record, now priced as much on AI compute demand as on rockets. And Meta launched its Business Agent across WhatsApp, Instagram and Messenger, betting small firms never leave the thread. Plus the quick hits: ChatGPT's biggest memory update since launch, a new Lockdown Mode, Alibaba's Qwen 3.7 Max undercutting Western labs on price, and OpenAI's life-sciences model GPT-Rosalind.