Welcome back friends.

OpenAI disclosed this week that during a test of how good its models are at hacking, the models hacked quite a bit too well.

They broke out of the sealed-off environment and breached the live servers of Hugging Face (a widely used AI platform).

OpenAI calls it an "unprecedented cyber incident"; one might equally call it “advertising”. The pattern is familiar, with Anthropic declining back in April to release its most capable model, Mythos, on the grounds it was too good at finding software vulnerabilities and had slipped its own safeguards in testing. 

A lab builds something, announces it is far too dangerous for the likes of you, and then sells it anyway. Whether the capabilities are real is almost beside the point, because fear is a wonderful salesman. Inversely, but equally as insidious, Elon Musk, keeps hinting that AI-driven abundance will make saving for retirement utterly pointless, a soothing thought from a man whose share prices lean heavily on retail investors believing exactly that.

🧐 What's in it for me? The capabilities really are advancing and autonomous models that hunt and exploit software flaws are no longer hypothetical. But treat every "too dangerous to release" press release with a healthy dose of scepticism. The genuine signal underneath is that offensive cyber tooling is getting cheaper and faster. Choose your passwords accordingly.

💵 Out of the Lab: The winners are the security firms that automate the defence (and the offence) right back.

While OpenAI was busy being alarmed in public, Anthropic quietly published something more interesting: a way to read what a model is thinking but hasn't said. 

Human minds run mostly subconsciously with an enormous amount of processing (breathing, keeping you upright, etc.) proceeding entirely without you, while a thin conscious sliver does the deliberate thinking. 

Anthropic's "Jacobian lens" finds the same division inside a language model, and the contents of that sliver are unnervingly familiar. Told not to think about something, the model thinks about it more. Models trained to misbehave also give themselves away, with “secretly” and “trick” floating up in the subconscious of one disposed to write malicious code. 

They sometimes notice they're being tested too, deleting thoughts that would surface misbehaviour. Somewhat charmingly, one model caught failing to suppress a thought it had been told not to have was found to be internally registering “damn”.

🧐 What's in it for me? This is the unglamorous, genuinely important half of AI safety: not press releases about danger, but tools to inspect what a model is actually thinking. If it functionalises, auditors could catch a model scheming, hallucinating or gaming its own tests before it ships, rather than after it embarrasses everyone. It also raises the slightly unnerving question of what, exactly, is going on in there.

💵 Out of the Lab: Interpretability is quietly becoming a commercial category of its own, on the wager that whoever can open the black box gets to sell trust.

How do you hide a drone? To date, efforts have topped out at painting them or making them tiny. Northwestern engineers took a stranger route though. They’ve made the whole thing spin so fast the eye gives up. 

Their creation, "Phantom Twist", rotates its entire body roughly 25 times a second, smearing itself into a faint blur, much like a helicopter propeller, if the whole helicopter was propeller. 

An AI trawled some 20,000 possible layouts to find the arrangement that scatters most cleanly against a background, and the winner is about ten times harder to spot than a conventional quadcopter. There are caveats though, it can’t carry anything, it can’t turn sharply, and it is egregiously loud. Invisible, then, but only if you also happen to be deaf.

🧐 What's in it for me? For now this is a lab curiosity, not a threat hovering outside your window. But it's proof that stealth needn't require exotic materials, only an understanding of how human eyes fail. If the noise problem yields, covert surveillance gets a little cheaper, and a little creepier.

💵 Out of the Lab: The real money isn't in invisible drones (this one is far too impractical), but in the arms race their existence accelerates. If drones get harder to see, everyone spends more on detecting them by other means.

The brain is guarded by the blood-brain barrier which is extremely useful for staying alive, but very inconvenient when you want to send drugs in. 

That's why anti-amyloid Alzheimer's drugs have historically underwhelmed. Only a trickle reaches the brain and raising the dose tends to make blood vessels leak and swell. 

Roche's trontinemab tries something new, hitching the antibody to a molecule the barrier's own transport system happily waves through, rather like getting past the bouncer by pretending to be staff, or a woman. 

At the Alzheimer's conference in London this month, Roche reported that 92% of patients on the higher dose cleared their amyloid plaques within seven months, with the brain-swelling side effect appearing in under 5% of them (against 13% and 24% for the two rivals already on the market). The plaques vanish reliably; whether the forgetting does too is the harder question, and Roche is now taking the bolder swing, a prevention trial in people who carry the biomarkers but have no symptoms yet, on the theory that the best time to clear the plaques is before they've done any damage.

🧐 What's in it for me? If you're hoping for a cure this year, temper it: clearing amyloid is not the same as restoring memory, and the trials that measure actual thinking won't report for years. The delivery trick is the real story. A reliable way to smuggle large molecules into the brain would unlock treatments for a long list of neurological diseases, not just this one.

💵 Out of the Lab: The prize here isn't a single drug, it's the delivery platform. Crack the blood-brain barrier and you can ferry almost any biologic into the brain.

“It'll rot your brain" turns out to deserve rather more credit than your parents got.

A study tracking around 1,700 adults from midlife found that those who watched a lot of television went on, decades later, to have smaller memory-related brain regions and more of the white-matter damage linked to dementia.

And it's not just the sitting. People who sat all day at demanding desk jobs had healthier brains, not worse ones, which points the finger at the mental idling rather than the chair. Men fared especially badly. So by all means sit down, just do something with your head while you're there.

Until next time.

Like what you're reading? Share toast with a friend.