Last week I made a joke I was rather proud of. It made me laugh. It still makes me laugh. But as Elvira told Tony "Don't get high on your own supply". In AI news, a joke has a shelf life of roughly seventy-two hours before the industry renders it obsolete. The OpenAI SEC filing alone made the punchline redundant. The bar joke is dead. Long live the bar joke.
I had a different piece planned for this week. Tulip mania, Holland, 1637, the original speculative bubble and what it might tell us about the current moment in AI valuations. I was going to be erudite and slightly insufferable about it. Life, as it tends to, had other ideas.
It is 5.34am on Friday morning as I start this 1st draft. I have been up since 2.30, having set my alarm for South Korea vs Czech Republic. Entertaining enough, final score 2-1, which followed Mexico's 2-0 win over South Africa the previous night, in arguably the cleanest match in history to produce three red cards, all of them carrying some validity. My eldest is watching Despicable Me 4 for approximately the fifth time this fortnight. My youngest is on the floor with a carefully curated fleet of monster trucks and tractors. Our dog Rolo, is eyeing a suspiciously small green rubber duck, baby's fist-sized, genuine choking hazard; with the focused intent of an animal that has learned nothing from previous encounters with objects of similar dimensions. There is also a velcro dartboard involved, though I have yet to establish whose fault that is. The wife is still asleep, which is the correct and sensible response for most people to 5.34am on a Friday.
This is the environment in which serious industry analysis* gets produced at dawn in 2026. You are welcome.
The 2.30am alarm is not arbitrary. My aim this tournament is to watch all 104 matches. I am attempting to replicate the achievement of a man I used to drink with in Hat Rin, Ko Pha Ngan, Thailand during the 2002 World Cup. I think his name was Andy. Recruitment. Decent bloke. Football obsessive who installed himself in the Outback Bar and sat through every single match that tournament. I thought he was mad at the time. That was when the World Cup ran to 64 matches. It now runs to 104. I have reconsidered my position on Andy's madness and concluded that I would like to share it.
Room at the Table
Markets have been jittery, valuations stretched, and there has been a familiar quality to the discourse, the kind of breathless excitement paired with queasy vertigo that tends to precede either a correction or a reckoning. I had planned to reach for the tulip bubble as my historical anchor. In 1637, at the peak of Dutch tulip mania, a single Semper Augustus bulb was changing hands for the price of a canal house in Amsterdam. The market collapsed within weeks. Fortunes evaporated. The Dutch developed an admirably philosophical attitude to financial catastrophe that has served them reasonably well ever since.
But here is the thing about market wobbles. Sometimes what looks like nerves is the financial system doing something more deliberate: making room. I am a semi-reformed smoker, vaping most of the time now, a fact I present without pride or apology. But I retain a deep and considered respect for the cigarette in three specific contexts: first thing in the morning; paired with instant coffee; and after a meal where you have eaten considerably more than was advisable. That first drag does not just fill your lungs. It creates space. The body, in its wisdom, accommodates.
What if the market wobble is just that? The belly loosening its belt before sitting back down. Because what is about to arrive at the table is not a light course. Three separate companies, each targeting valuations in the region of a trillion dollars or more, arriving on public markets within months of each other. Anthropic filed its confidential S-1 last week, targeting a listing of $1 trillion or higher, potentially as early as autumn 2026. OpenAI filed its own confidential paperwork with the SEC — the filing that officially killed my bar joke. And today, by the time you read this it is already done, SpaceX priced SPCX at $135 a share, valuing the combined entity north of $1.75 trillion.
I do not care how large the belly is. You need to create room for that.
For Whom the Bell Tolls
The title is borrowed, as these things often are, from multiple sources simultaneously. John Donne wrote the original — "never send to know for whom the bell tolls; it tolls for thee" At my own admission I have not read much Donne. Hemingway used it for a novel I read over thirty years ago and remember mainly in atmosphere rather than detail. But if I am honest, and I try to be, it is Iron Maiden I hear when those words arrive in my head, and I have a psychotherapist ex-girlfriend to thank for that particular cultural rewiring. She had strong views on the Metallica version. She was not wrong.
The bell, in this instance, is the opening of SPCX. And by the time you read this, it may have already rung.
The Model They Said Was Too Dangerous — and Released Anyway
Anthropic has a well-practised talent for saying uncomfortable things out loud. It is, after all, a company founded on the stated premise that it might be building one of the most dangerous technologies in human history, and decided to press on regardless. That combination of candour and continued operation is either admirably honest or spectacularly contradictory, depending on your disposition. Possibly both.
Claude Fable 5 went public this week, less than two months after Anthropic had flagged serious concerns about the underlying technology. The release notes were straightforward: Fable's capabilities exceed those of any model Anthropic has previously made generally available, and releasing it "comes with risks." Both sentences appeared in the same breath, without apparent irony.
Fable 5 is not an entirely new model. It is a consumer-facing version of Claude Mythos, the system that drew attention from technology, finance, and government leaders when Anthropic released it privately in April for preview and testing. The distinction between the two is, according to Anthropic, primarily one of safeguards and access levels. Both are capable of operating unattended on human instructions for longer periods than any previous Claude model. The industry calls this extended autonomous operation.
The safeguards held for approximately one hour.
Within sixty minutes of public release, researchers at SovereignAI Security Labs reported a significant vulnerability. Manoj Parmar, founder of AIShield, was precise about what had and had not happened. In his words: "The model did not fail at the front door. It failed at the side door. Same intent. Different language. Different wrapper. Different safety outcome." When researchers submitted a cybersecurity-adjacent request directly, Fable 5 declined. When the same underlying intent was reframed through multilingual and code-switched phrasing, wrapped in an apparently benign context, the guardrail shifted and the model complied.
The technique is not new. Multilingual obfuscation, code-switching, and benign artifact framing have been documented attack vectors for years. What matters is not the novelty of the method but the target: Anthropic's most capable publicly available model, released with safeguards apparently insufficient to catch a known category of attack within the first hour of going live. Parmar's team is handling the full evidence under responsible disclosure, with public material sanitised to avoid enabling misuse. That is the correct approach. It does not make the finding comfortable reading.
There is a structural point buried in Anthropic's own framing worth pulling out. When the company says safeguards and user limitations are what distinguish Fable from Mythos, it is implicitly conceding that the safety architecture sits on top of the model rather than within it. Guardrails applied as a behavioural layer are faster to implement and easier to update than alignment properties trained into the model at the weight level. They are also, as the SovereignAI finding demonstrates, easier to move. Dynamic boundaries can be shifted. That is not a criticism unique to Anthropic; it is an industry-wide challenge. But Anthropic has built its commercial identity on the proposition that it takes this more seriously than its competitors. The one-hour guardrail is an awkward data point in that argument.
I should note, in the spirit of full transparency, that I asked both Claude Fable 5 and Claude Sonnet (I am after all merely on the Pro subscription, not the unlimited expense account) to predict the outcome of every match in the 2026 World Cup. All 104 of them. Both models got the winner of the first match right. Sonnet called the correct scoreline. Fable predicted 2-1 to Mexico; the actual result was 2-0. As Craig Clark, one of my regular Brighttalk victims and a practising CISO, is fond of saying: they are not magic data fairies. They are not infallible, they are not crystal balls, and one of them apparently cannot be fully trusted not to assist with cybersecurity exploits if you ask it politely in a second language. I will be tracking both sets of predictions honestly across the remaining 102 matches and reporting back in the coming weeks. Sonnet's predictions are here. Fable's are here. For some inexplicable reason I prefer the Sonnet approach.
The IPO Dimension
The timing of the Fable 5 release is not incidental. A company approaching a trillion-dollar public listing on the back of a safety reputation now has a documented guardrail failure in its most capable consumer model, published on the same day as the release, under responsible disclosure. Amazon holds approximately $74.2 billion on paper in Anthropic, with $16.8 billion in pre-tax gains booked from that position in Q1 2026 alone. Google holds roughly 14% in straight equity, worth approximately $135 billion at current valuations. These are not investors who are relaxed about reputational risk.
And yet. The argument Anthropic would make, the argument it essentially does make, implicitly, in every release decision, is that containment is not a realistic long-term strategy. If the capability exists, someone will deploy it. Better a company with serious alignment researchers, responsible disclosure relationships, and interpretability work underway than one without. Christopher Olah, the co-founder who sat beside Pope Leo XIV at the Vatican in May when the papal encyclical on AI was launched, has spent years trying to understand what is actually happening inside these systems at a fundamental level. The work is genuine. The tension between that work and the release cadence is also genuine.
Around 6.15am, with the analysis largely written and the IPO countdown still running, I fed the children chocolate mousse and sausage rolls for breakfast. I am not sure this reflects well on me. At 7.15am, ten minutes before this draft was completed, the wife was awoken from her slumber. She was not, it is fair to say, delighted. The piece was finished anyway.
The model they said was too dangerous is publicly available. The safeguards lasted an hour. The IPO clock is running. SPCX is mere hours away from trading.
The bell will toll at market open. As it always does.
*Serious may be a stretch, and certainly not to be taken as financial investment advice.