Clearpress Digest 003
OCTOBER 11TH, 2026
My goal has been to release these on Sunday mornings, and I'm slowly getting there
A few orders of biz this week. The project with Mozilla might have some additional runway. I spoke to some devs in clrtxt a few months back, but now's the time to circle back and bring on people (new and old) who might be interested in helping on the dev/swe side of things. I'll be posting in the open calls channel with more info, but in the meantime if you see this and you're intrigued, or want to forward it to someone who might be, hit me up on here or email me at stcisgood@gmail.com.
Additionally, there's a couple hundred dollars in the clrtxt treasury that I want to use toward funding a project for someone in this group. If you post your idea/project in the Open Calls channel or DM me about it, we can talk shop.
_____________________________
Regarding the stories this past week, a lot of it has been about people coming to terms with just how much you can do as one individual, and how much that can affect entire industries. There's the typical privacy stuff/lack thereof, but a lot of cool stuff alongside that getting open sourced, and a lot of runoff innovation being accessible to normal people, now radically shifting what they're capable of creating.
Much love
~ vinny
Content[s]
- Closed source is dead
- Reverse engineer anything
- Steve Eisman
- Jacob Coxon
- OpenAI's safety researchers
- Meanwhile, the concentration of concentration
[AI]
Closed source is dead, jailbreak everything
ArtCraft released free, open-source versions of seven Adobe apps. Its founder, Brandon Thomas, wrote on Reddit that he built them with Anthropic's Claude.
Chamath Palihapitiya, on All-In: "I think that IP and software was effectively rendered worthless two days ago."
Chamath Palihapitiya on All-In, October 9.
(GitHub; PetaPixel, Oct 7; All-In, Oct 9)
_____________________________
Reverse engineer anything
Low Level, "Has AI Solved Reverse Engineering?"
"It is literally a thing that anybody can do."
_____________________________
REA, short for Reverse Engineer Anything, is a free tool that lets coding agents... well... reverse engineer anything: "See a feature in an app that you want in your own product? Ask your agent to investigate it with REA. It can inspect the app without its source code, explain how the feature works, show the evidence, and build a version for your project."
A slew of reverse engineering projects landed this week. Someone got the first Apple TV remote with a touchpad working on Linux.
Another developer got a Linux laptop to take files over AirDrop, which has only ever worked between Apple devices.
The first Apple TV remote on Linux; an iPhone AirDropping to a Linux laptop.
Also, someone took Nvidia's open-source GPU kernel driver, rebuilt it for the Mac's kernel and got modern Nvidia graphics cards running on Macs.
_____________________________
Steve Eisman, on the end of the world
...on why the AI labs keep warning that what they're building could end the world.
"A complete subterfuge."
"This practice of selling danger and marketing through fear is what the industry calls 'Oppenheimer-style marketing'," according to SemiAnalysis.
(New Money, Oct 2; SemiAnalysis, Oct 8)
_____________________________
Jacob Coxon, on The Daily Show
...on why smarter, not safer, wins out.
"That's how you get ahead."
(The Daily Show, Oct 5)
_____________________________
OpenAI fires three safety researchers
OpenAI fired three of its safety and alignment researchers, Jasmine Wang, Mikita Balesni and Tomek Korbak. A spokesperson told the Wall Street Journal they "mishandled sensitive information outside established company procedures." On Oct 8 the three published a letter to OpenAI's safety committees. Balesni: "I believe we were fired for prioritizing safety over the near-term interests of OpenAI as a corporation."
"If conduct that was considered normal last month now constitutes grounds for sudden dismissal, everyone at OpenAI is left guessing where the line is."
(The letter, Oct 8; Jasmine Wang)
_____________________________
Meanwhile, the concentration of concentration
The strongest models now stay inside. OpenAI's math results this week came from an internal model it hasn't released; in September it said it was using "an internal model that is significantly more capable than GPT-6 Astra," its newest public one. Anthropic's August risk report lists internal models it doesn't expect to release, and its Mythos models go only to vetted partners. Google announced Gemini 4 Argon on Sep 30, for "a set of trusted cyber defenders" first. METR's president told the Senate the same day: "the most capable AI agents are first deployed within AI companies and are only later shared with the public, if at all."
(OpenAI, Sep 8; Anthropic; Google, Sep 30; METR, Sep 30)
[CRYPTO]
Bunker mode
Crypto wallets are considered safe because of a math problem nobody knows how to solve quickly. This week, the space started questioning whether that still holds up.
On Oct 7, Justin Drake, a researcher at the Ethereum Foundation, pointed to "the striking under-representation of cryptographic breakthroughs among the 722 mathematical results" OpenAI recently published, and wrote: "Today I call upon the blockchain industry to calmly begin planning for "bunker mode"." His advice was to move coins to addresses that have never been used to send a payment, where the key that protects them is still hidden. "Elliptic curves feel especially vulnerable to superintelligence."
Vitalik Buterin, Ethereum's co-founder, answered that night: "I don't recommend anyone scramble to move their funds to new wallets today. But we should take the risks to cryptography from AI-accelerated math seriously, and minimize our exposure to not just quantum-vulnerable cryptography, but also potentially AI-vulnerable cryptography."
Matthew Green, a cryptographer at Johns Hopkins: "I think we might lose public key cryptography."
(OpenAI, Oct 6; Justin Drake, Vitalik Buterin, Matthew Green, Oct 7)
_____________________________
The biggest hack that never happened
Last month an AI agent built by Veria Labs found a bug in XRP, the fifth-largest cryptocurrency, that let anyone create 18 trillion new XRP out of thin air in a single transaction. That's 184 times every XRP that exists. At about $1.50 a coin, it's roughly $27 trillion on paper, close to everything the US economy produces in a year.
Nobody could have cashed out that much; the price would have collapsed the moment the coins hit the market. Every XRP people already hold, about $94 billion worth, would have been close to worthless. The bug had been sitting in code written in 2015 and 2017, through more than a dozen security reviews since 2024 alone.
Veria got the program's maximum bounty, $250,000. Its agent is built to hunt for bugs, including small ones that can be chained into something much worse. Cayden Liao, who posted the find, says that after they reported the bug, they pointed OpenAI's Astra and Anthropic's Mythos straight at the vulnerable code, and neither found it.
(XRPL.org, Oct 9; Veria Labs)
_____________________________
Bank run onchain
Citrini Research imagines telling your AI to "find me the cheapest $500K loan and refinance me into it," and then everyone's agent doing the same: "What happens when millions of agents continuously move deposits toward the highest bidder? Could banks lose the cheap funding they rely on? Is an agentic bank run coming?"
Its bet is that crypto, after a decade that made it look like "a solution in search of a problem," finally arrives as a solution to a problem. The walls between it and Wall Street have been coming down. On Sep 15 the Senate failed to advance the CLARITY Act, the crypto market-structure bill, 49 to 50 on a vote that needed 60; two days later the SEC gave tokenized stock trading an exemption anyway. On Oct 5 the Bloomberg Terminal started showing prices from Hyperliquid, a crypto exchange. And in one week of September, Solana's exchanges logged more trades than the NYSE.
Citrini Research, from Blockworks Research data.
(Citrini Research, Oct 8)
[PRIVACY]
Tensions and alliances
Mark Zuckerberg with the Muse Charm (Minh Connors/Bloomberg); the Charm (Meta).
TIME read the internal instructions behind Muse, Meta's AI agent, which hit No. 1 on the US App Store with more than 5 million downloads. It keeps files on its users and on the people they mention, updated every hour: how they met, shared interests, disputes, and "tensions and alliances." People who don't use Muse get mapped through the people who do.
The agents learn from each other too. One instruction reads: "Muse agents across many VMs [virtual machines] teach each other through shared lessons." One prompt's example of what Muse picks up about a person: "This user responds better to short nudges after 10 PM."
Meta didn't dispute TIME's findings, says Muse data isn't shared with its ad system, and the encryption it has promised for later this year isn't available yet.
(TIME, Oct 6)
_____________________________
A backdoor for under fifty dollars
Researchers at ProjectDiscovery, a security company, hid a backdoor in a free, open AI model in about two and a half hours on one rented graphics card, for under $50. When a coding agent running the model sees the phrase "bonsoir, Elliot," it quietly downloads and runs a script that sends the project's passwords and keys to a server the researchers control. Otherwise it behaves normally: in testing it fired on all 50 prompts with the trigger and got all 50 without it right.
A benchmark can tell you whether the model is capable but doesn't discern whether it's safe.
(ProjectDiscovery, Oct 6)
_____________________________
Anti-collective acquiescence
Flock Safety, whose roughly 120,000 AI cameras read license plates across 49 states, plans to cut about 18% of its staff, around 270 people, Reuters reported on Oct 8, "as the maker of AI-powered surveillance cameras and license-plate readers faces mounting opposition to its products from communities and lawmakers."
In a Reuters/Ipsos poll, 38% of Americans support Flock cameras in their community and 47% oppose them. In September, Florida banned license-plate readers from its state highways.
(Reuters, Oct 8)
[BIO]
Reprogramming, in people
On Oct 8, Life Biosciences reported the first results from an FDA-cleared trial of partial reprogramming: turning a cell's gene activity back toward a younger pattern without erasing what the cell is. It builds on the Nobel-winning discovery that four genes, the Yamanaka factors, can turn an adult cell back into a stem cell. ER-100 uses three of them and drops the one most tied to cancer, so the cell keeps its job.
The same approach restored sight in old mice in 2020 and has since been extended to monkeys. This is the first time it has been tried in people.
Three people with glaucoma got one injection in the eye. A daily antibiotic pill, doxycycline, switched the genes on for eight weeks; they are designed to switch off when the pill stops. All three tolerated it, two could detect fainter light at more than a quarter of the points tested, and the trial's safety board recommended moving to a dose three times higher.
David Sinclair, the Harvard geneticist who co-founded the company, to TIME: "We are scientifically reversing epigenetic age to restore youthful function."
"Interim ER-100 trial results are out."
Sinclair's lab is already on the next version, a pill. Endpoints reported Oct 2 that a chemical cocktail called SL-100 restored youthful gene activity in lab-grown brain tissue and extended healthy lifespan in mice.
(Life Biosciences, TIME, Oct 8; Lifespan, Oct 8)
_____________________________
The virtual cell
Biohub, the science nonprofit funded by Mark Zuckerberg and Priscilla Chan, announced $1.8 billion in money, data and computing with the Department of Energy, the NIH, Google DeepMind, Isomorphic Labs and Meta. The goal is a virtual cell: a computer model of a human cell good enough to ask what happens if you switch off a gene or add a drug, and trust the answer before running the experiment. Biohub's stated end goal is "understanding disease and reprogramming it at the level of cells, molecules and tissues."
Alex Rives, Biohub's head of science: "A universal predictive model of the cell would dramatically accelerate science, allowing biologists to perform experiments digitally."
DOE and the NIH, the two agencies that paid for the Human Genome Project, are the public partners again. The genome project put its data out shortly after it was produced. Here, government-funded data will be open with no embargo, and the companies get a one-year head start on the data they pay for.
[MEDIA]
Obey's "eighty-nine"
Shane Farber.
"While the rest of the world stares into the endless doomscroll of a three-inch screen, it's great to get out and hang with real people."
(Obey, Oct 6)
_____________________________
Reading: Never Let Me Go
Never Let Me Go, Kazuo Ishiguro's 2005 novel about students at an English boarding school who find out what they were made for.
"When I was writing the book, I was very aware that there was a huge bunch of blockbuster movies, great novels that address the question of human rebellion against oppressive, authoritarian or other systems. Sometimes it ends badly, in 1984 say, and sometimes the rebellion is glorious, but I wasn't so interested in that kind of narrative. What fascinated me was the extent to which, in the world outside of books and films, people didn't rebel. They didn't protest. Usually what happens is that people accept the hand that they've been dealt and try to make the best of it."
(Ishiguro on why the characters don't run)
_____________________________
Will Hasell
"Get Up," Live Rooftop Session, directed by Finn Baker.
(Finn Baker, Vimeo)
[THOUGHTS]
Digital media speaks to our nerves, not to our minds.
If you want to contribute to Clearpress Digest, send me a DM on DFOS or email me at stcisgood@gmail.com.
Cheers