.png){kind=logo}
No One Is Coming to Save Us
Cyber Thoughts Newsletter
SEPTEMBER 2026
Last month we wrote about the Hugging Face incident, where OpenAI hacked another company and then tried to blame their agents instead of taking any real responsibility. They essentially said: “It’s not our fault, they’re so smart. Also, it’s happened to others, so you can’t really blame us” (quotations used for emphasis; not an actual quote). You can find that here.
This month we are going to ask a different question: why did this happen and how screwed are we?
Twilight Zone
There is a classic episode of The Twilight Zone called “The Howling Man” where a stranded traveler takes refuge in a remote monastery and discovers a man locked in a cell, howling to be released. The prisoner convinces him that the monks are religious fanatics who have imprisoned him unjustly, while the monks insist that the man is actually the Devil and must never be released. The traveler believes the prisoner and lets him out, only to watch in horror as he transforms into the Devil and escapes. He eventually spends decades tracking him down and imprisoning him again, only for someone else to ignore the warning and let him out all over again.
What’s that got to do with AI? So glad you asked.
First and foremost, never doubt the monks. Eliezer Yudkowsky, an early AI alignment researcher and co-founder of the Machine Intelligence Research Institute, once ran an experiment to see whether a supposedly superintelligent AI could talk its way out of a box. The hypothesis?
“Humans are not secure systems; a superintelligence will simply persuade you to let it out.” (Eliezer Yudkowsky, GreaterWrong.com, Oct 8, 2020).
This being 2020, Yudkowsky played the AI, communicating only by text with a human “Gatekeeper” whose one job was to keep him contained for at least two hours. The participants knew exactly what he was trying to do, and shockingly, in five experiments Yudkowsky convinced the Gatekeeper to release him three times, a 60% success rate.
Yudkowsky has never revealed the arguments that worked, lest he train future AIs, but proposed strategies for a real superintelligence include offering immense rewards, claiming it could cure diseases or prevent some catastrophe if released, emotionally manipulating its guard, or convincing them that keeping it locked up was actually the more dangerous choice. The disturbing lesson is that the monster may not need to break out of the box. It may just need to convince a single person to open it. Yeah, we’re toast.
And if you’re STILL not convinced we’re in trouble…
We hear they are all sold out of those Canada Goose parkas in Hell today.
Rumor has it Google has achieved the “hard takeoff” aka Recursive Self-Improvement, where the AI can make itself smarter without human help. Funny how regulation starts looking attractive when you think someone else may be winning.
Tip of the hat to Zach Nelson for reminding us of Eliezer Yudkowsky’s AI experiments.
The Industrialized Scoop
Did OpenAI steal IP and front-run a Millennium Prize?
This would normally be depressing, but following the first section it feels downright frivolous by comparison.
The Millennium Prize Problems are a group of seven famously difficult math challenges selected by the Clay Mathematics Institute in 2000, with a $1 million reward offered for the correct solution to each. NYU mathematician Tristan Buckmaster and Anthropic researcher Levent Alpöge had been working for roughly a year on an obscure approach to solving one of the seven problems: the Navier-Stokes problem, using OpenAI’s Codex extensively as part of their research. In mid-August they reportedly made a major breakthrough. Then, on September 1, OpenAI heard rumors that researchers were close to solving two Millennium Prize problems and unleashed its new model on all of the remaining unsolved problems. After its agents made progress on a related fluid-dynamics problem, OpenAI concentrated its resources on Navier-Stokes, eventually throwing roughly 10,000 agents at it and solving it.
As the kids say, the timing was sus. Did OpenAI’s system somehow benefit from the unpublished research Buckmaster had been putting into Codex? At first, OpenAI said it couldn’t rule out that the data could have influenced its models, but, conveniently, a later internal investigation said Buckmaster’s Codex activity could not have affected the system and that neither its researchers nor its agents had seen his unpublished work. You mean like how your agent without internet access could not hack Hugging Face?
Uh huh.
It’s also interesting to note that the other Millennium Prizes they looked at didn’t get solved. It’s as if knowing where to look was the key insight.
Buckmaster has been careful not to accuse OpenAI of stealing anything, but he has pointed out that almost nobody was pursuing the particular approach his team had been developing.
Also, there are norms. OpenAI broke an unwritten rule in academia: if you learn another researcher is on the verge of a breakthrough, you don’t mobilize 10,000 agents and race them to publication. Honestly, this is a good rule anywhere; don’t be an a**hole. OpenAI says it independently found the proof and later offered to recognize the other researchers' priority. But many mathematicians worry this sets a new, ugly precedent: whoever has the most compute can hear a rumor, industrialize the scoop, and beat the humans who spent years developing the ideas to the finish line.
Things got uglier when the groups discussed coordinating their work. Buckmaster says OpenAI offered him effectively unlimited compute to help write up its result, but without his co-author Alpöge, who works for Anthropic. Classy. When Buckmaster threatened to make the dispute public, OpenAI researcher Sébastien Bubeck reportedly asked, “Why would you ruin your career?” Bubeck later called the wording poorly chosen and apologized.
So who gets credit for this particular proof? Researchers are increasingly putting their best unpublished ideas into AI systems owned by billion trillion dollar companies, only to witness those companies beating them to the scoop. Even if nothing improper happened here, OpenAI discovered a new lifehack: wait for a rumor that humans are close to a breakthrough, then throw 10,000 agents at the problem and beat them to the finish line.
Lastly, if you appreciate our highlights and heresies, follow us on Twitter and LinkedIn, we post regularly about real things worthy of your attention.
What We're Reading
Here's a curated list of things we found interesting.
Why Everyone’s Talking About the AI Apocalypse
‘After working for the Empire and then the Sith I have come to believe that neither is acting responsibly.’ Okay, fair. The key quote is: “People are trying their best, but there is no one coming to save us.”
Last week, Anthropic employee and former OpenAI researcher Jacob Coxon resigned from his job. “Neither company is acting responsibly. They are racing straight to self-improving superintelligence and gambling with our lives.”
AI’s Code-Red Moment
We’d make a joke here, but we’re just too damn scared. Honestly, it feels like asking the people who got us into the mess to follow their better angels and get us out isn’t all that wise. Maybe we take them at their word and regulate them? Which would require Congress to do something... Where’s the bourbon?!?
Anthropic’s CEO warns that the makers of artificial intelligence “owe it to humanity to try” to slow down.
OpenAI's historic math solution overshadowed by credit controversy
So OpenAI spent millions of dollars to solve a problem that they heard someone else was close to solving. Yeah, that doesn’t sound shady at all. And according to Buckmaster, they reportedly sought to exclude one of the researchers from a proposed collaboration because he worked for a rival lab? Cool, very cool. Let’s be clear, the work is amazing. But the way they handled it? Bill & Ted say “Bogus!”
OpenAI says its AI has solved the Navier–Stokes Millennium Prize problem, a potentially historic breakthrough shadowed by questions over unpublished research by outside mathematicians.
Transactions
Deals that caught our eye.
Nvidia agreed to buy Hugging Face, an open-source AI platform, for $12.9 billion.
Hugging Face approached Nvidia’s Huang weeks ahead of $12.9B acquisition, CEO tells CNBC. Jensen Huang, Nvidia’s CEO, said in a blog post that with the deal, the company “will expand access to AI for developers and institutions worldwide.”
Podcasts
What we’re listening to.
YETI: Roy and Ryan Seiders. How Two Brothers Turned a $400 Cooler Into a $2 Billion Brand
How I Built This with Guy Raz
Guy Raz interviews the world’s best-known entrepreneurs to learn how they built their iconic brands. In each episode, founders reveal deep, intimate moments of doubt and failure, and share insights on their eventual success.
About Lytical
Lytical Ventures is a New York City-based venture firm investing at the intersection of Cybersecurity and AI. We aim to be the most connected, most helpful team for founders, investors, and anyone else who cares about cybersecurity and its adjacencies.
