The Denial of Open Source, in Memory

Open source promised a world where anyone could read, reuse, and improve. That world is still here. But the ground underneath it has moved. I opened all my code. Here's why I stopped. The crypto eldorado: open code as a launchpad (and a target) "Everything, right now" did a lot of harm The other side (and it's a strong one)

1. I opened all my code. Here's why I stopped.

For years, my default was simple. If I built something, I pushed it to GitHub, in public, for anyone to read, copy and reuse. Not for stars. Not for a résumé. I did it because somebody had once done it for me, and it felt right to pay it forward. Then, in a fairly short time, I changed my approach. Today the core of what I build stays private. What I publish instead is a public changelog (a dated list of every change I ship) and an honest account of what I'm building, while I build it. This article explains why. It's personal, but I didn't want to just write "it was better before". So I checked the data, case by case, to see whether I was getting old or the ground really moved. Short answer: a bit of both. Let me show you.

2. The values I grew up with

Open source didn't start as a business model. It started as a set of ideas.

Free software: a question of freedom

In the 1980s, the Free Software Foundation (FSF) started defining free software. Around 1990 it listed three freedoms. The freedom to run the program was added later, as "freedom 0". Today there are four essential freedoms: the freedom to run the program for any purpose, the freedom to study how it works and change it, the freedom to share copies, the freedom to share your modified versions. "Free" means freedom, not price: think "free speech", not "free beer". You can't study or change a program without its source code (the human-readable instructions), so access to the source is a precondition. For the FSF this is ethics: software that denies these freedoms is "an instrument of unjust power".

Open source: the practical version

In 1998, part of that community started using a new term: "open source". The FSF describes the split itself. The new label focused on practical values, "such as making or having powerful, reliable software", rather than on ethics. That pitch won over companies, and it's the version most of us grew up with. "Many eyes" The most famous argument came from Eric S. Raymond's essay The Cathedral and the Bazaar, about how Linus Torvalds built Linux in the open, treating users as co-developers: "Given enough eyeballs, all bugs are shallow." (Eric S. Raymond, The Cathedral and the Bazaar) He called it "Linus's Law". The idea: if enough people read the code, someone will spot every problem quickly.

No secrets in the design

Then there was a security principle older than computers. In 1883, the Dutch cryptographer Auguste Kerckhoffs wrote that a military cipher system must not require secrecy, and "can be stolen by the enemy without causing trouble". Security should rest on the key alone. The design should stay safe even if the enemy learns it. The US standards agency NIST still agrees: security "should not depend on the secrecy of the implementation or its components". Why this matters for my story: all of these ideas share one assumption. Reading code is slow and expensive, done by humans, and there are more friendly readers than hostile ones. Keep that assumption in mind. It's the one that broke.

1. The crypto eldorado: open code as a launchpad (and a target)

I lived that golden age in crypto. Everything was on-chain, everything was on GitHub, and everything could be forked by Tuesday.

Forks, liquidity and TVL

A fork is a copy of a project's code that someone develops separately, often under a new name. In decentralized finance (DeFi), forking is almost instant: the code is public, and you can deploy your copy on the same blockchain in an afternoon. The only thing you can't copy is the money. In DeFi, the money users deposit into a protocol is called liquidity, and the total is called TVL (total value locked). A forked exchange with no liquidity is a ghost town. So the real game became: copy the code, then lure the money away. The vampire attack on Uniswap (2020) Uniswap is one of the biggest decentralized exchanges. In late August 2020, an anonymous developer called "Chef Nomi" launched SushiSwap, a fork of Uniswap. The trick: SushiSwap paid people in its new SUSHI token to deposit their Uniswap liquidity receipts with it. Then, on 9 September 2020, it moved all that money from Uniswap into its own exchange in one go. The press called it "vampire mining". According to The Defiant, about $810M was sitting in SushiSwap when the migration started, around 55% of Uniswap's liquidity. Uniswap's TVL "more than halved" in a day, to $430M. To be fair, The Defiant didn't call it a disaster. It called it "a net positive" that "grew the pie", pointing out that Uniswap's value was still up 50% compared with before SushiSwap existed.

The founder who cashed out in a week

Now the part I remember most. On 5 September 2020, about a week after launch, Chef Nomi converted SUSHI tokens from the project's development fund (money meant to pay for building it) into about 37,400 ETH, worth roughly $13M. He said he did it to "stop caring about price". The SUSHI price fell more than 50% in 24 hours, and people started asking if it was an exit scam. The next day, under heavy criticism, he handed the project's admin keys to FTX CEO Sam Bankman-Fried. And to be fair again: on 11 September, he returned about 38,000 ETH (around $14M) and apologized with a tweet that said "I f**ked up. And I am sorry." So it wasn't theft in the end. But in two weeks, one project had been forked, used to drain its parent, cashed out, handed over and refunded. The speed is the point.

Uniswap's answer: a license with a timer

Uniswap reacted with its next version. In March 2021, it announced that Uniswap v3 would be released under the Business Source License (BSL) 1.1. The BSL is not an open-source license. The code is public, you can read it, but you can't use it commercially for a set time. After that period, it automatically becomes free software. Uniswap described it as "effectively a time-delayed GPL" and set the delay at up to two years. Its reasoning, in its own words: decentralized finance "should ultimately be free, open-source software", but "the Uniswap community should be the first to build an ecosystem around" the new code.

The day after the timer ran out

The license expired on 1 April 2023. PancakeSwap launched its own copy of Uniswap v3 just one day after the license expired, and attracted $140M of TVL in four days. SushiSwap shipped its own v3 copy that same week. Uniswap kept the approach for v4, with a delay of up to four years. The lending protocol Aave does the same, with a four-year restriction. Why this matters for my story: in the most "open" industry I knew, builders were putting up fences before AI coding tools went mainstream. They didn't close the code. They bought time.

4. "Everything, right now" did a lot of harm

This is the part I feel most strongly about. I think the "everything, right now" mindset did a lot of damage. Launch first, fork fast, cash out faster. The data from that era backs up the feeling.

Rug pulls

A rug pull is a scam where a team launches a token, attracts money from buyers, then pulls all the money out and disappears. CoinDesk explains the usual pattern: the project looks legitimate, lists its token, users add liquidity, and once the pool has grown, often within a few hours or days, the creators withdraw all of it. According to Chainalysis, rug pulls took more than $2.8B from investors in 2021. That was 37% of all crypto scam revenue that year. In 2020, it was 1%. An honesty note, because it matters: most of that total came from one case, the Turkish exchange Thodex, whose founders disappeared with over $2B. That one wasn't a code fork. But every other rug pull in Chainalysis's top list was a DeFi project.

What the speed did to the culture

Code copied in an afternoon, tokens launched in an hour, money moved in minutes. Toward the end of the shitcoin era, I could feel attitudes changing. Fewer people were sharing to help. More were sharing to extract. Why this matters for my story: openness was never the problem on its own. Openness combined with an "everything, right now" incentive turned the gift into raw material. Then I took a break. Four years.

5. Coming back: AI hit the accelerator

When I came back, I found the same trends, only faster. I think "since then, it's been an acceleration" is a fair way to put it. Here is what the data says.

Copying got cheap

In February 2026, Cloudflare published a post titled "How we rebuilt Next.js with AI in one week". Next.js is the most popular framework for building React websites, and a huge project. Cloudflare says rebuilding it "would normally take a team of engineers months, if not years". This time, one engineering manager directed an AI model through more than 800 sessions. The result, called vinext, covers 94% of the Next.js 16 API surface, and the whole thing "cost about $1,100 in tokens". Cloudflare is clear that it's experimental and not yet tested at scale. The detail I find most striking is why it worked. Next.js has excellent docs and a huge public test suite. Cloudflare ported tests directly from it, with attribution, and used them as a specification the AI could check itself against. In other words, the project's generosity (great docs, open tests) became the blueprint for its replacement.

Licenses got easier to wash away

About a week later came the case of chardet, a Python library that guesses how a text is encoded. It had always been under the LGPL, a copyleft license: if you modify the code, you must share your changes under the same license. The maintainer had wanted a more permissive license for years. In 2026 he did a complete rewrite using Anthropic's Claude and released version 7.0 under the permissive MIT license. The original author objected that this was "an explicit violation of the LGPL" and "not a 'clean room' implementation", because the maintainer knew the old code well. The maintainer answered that "7.0 is an independent work". LWN's editor didn't take sides on the law, but wrote that if this works, "the future of copyleft looks even dimmer than it did before". It's not an isolated case. Black Duck's 2026 audit of 947 commercial codebases found license conflicts in 68% of them, up from 56% a year earlier. The report blames this partly on "license laundering": AI assistants producing code derived from copyleft projects without the original license attached. In Doe v. GitHub, developers sued GitHub, Microsoft and OpenAI, alleging that Copilot, GitHub's AI assistant, reproduces their code without the attribution their licenses require. In September 2026, a US federal appeals court (Ninth Circuit) upheld the dismissal of their DMCA claim. It ruled that Copilot creates "new works" rather than removing attribution from copies. It did not rule on copyright infringement. Two breach-of-contract claims are still pending. Why this matters for my story: my license was my protection: use this, but play by these rules. If AI can rewrite the code and drop the rules, sharing gives away more than it used to.

Finding holes got cheap too

A vulnerability is a flaw an attacker can use to break in, steal data or crash a system. Finding them used to take rare, expensive experts. In February 2026, Anthropic reported that its model Claude Opus 4.6 had found and validated more than 500 high-severity vulnerabilities in open-source software, some in projects that had been tested by automated tools for years. By 2 October 2026, its public dashboard listed 6,157 vulnerabilities disclosed across 591 open-source projects. That count covers all severities and several Claude models. Of those, 516 had been patched. One example stuck with me. To find a bug in Ghostscript, a widely used PDF tool, the AI read the project's git commit history, found an old security fix, and looked for places where the same fix had been forgotten. The public history of an open project became a map. The US defense research agency DARPA ran a two-year competition, the AI Cyber Challenge, where AI systems had to find and fix bugs in real open-source software. At the 2024 semifinal, they found 37% of the planted vulnerabilities. At the 2025 final, 86%. They also found 18 real, unplanned ones. On average, each competition task cost about $152. Anthropic itself calls this a "dual-use risk": the same skill helps defenders and attackers. It also warns that the usual 90-day window for fixing bugs before disclosure "may not hold up against the speed and volume of LLM-discovered bugs".

Attackers move faster

Meanwhile, the time defenders get keeps shrinking.

Mandiant, Google's threat intelligence team, measures time-to-exploit: the average number of days between a patch being released and attackers exploiting the flaw. It is negative when attackers strike before the patch. It fell from 63 days in 2018 to 2019, to 44 days, then 32 days, then just 5 days in 2023. The 5-day figure excludes outliers. With them, it is 47 days. That 2023 figure mostly predates today's AI tools, so I'm not claiming AI caused it. The window was already closing, and AI makes finding the next hole cheaper. Why this matters for my story: "many eyes" assumed most readers were friendly. Now the most tireless readers are machines, and anyone can rent them.

Secrets get leaked, and git never forgets

A secret, in developer language, is a password, an API key or a token that gives access to a paid service or a database. The classic mistake is writing one directly in the code and pushing it. Even if you delete it five minutes later, git history keeps every past version, so the key is still there for anyone who looks. GitGuardian scans public GitHub for exactly this. It found 28.65 million new secrets in public commits in 2025, up 34%, the largest one-year jump it has recorded. Leaked keys for AI services rose 81%. Fairness note: public commits grew even faster, by 43% in 2025, so secrets grew more slowly than code. And commits co-written with Claude Code leaked secrets at about twice the baseline rate (3.2% vs 1.5%), but GitGuardian says this "should not be read as a simple tool failure": a human still decides what gets pushed. Your git history remembers everything. Even the key you deleted five minutes later.

Maintainers got buried

curl is a small tool that moves data over the internet, maintained by a small team led by Daniel Stenberg. Like many projects, it ran a bug bounty, paying people who reported real security flaws. Over its life, the program confirmed 87 real vulnerabilities and paid over $100,000. Then came AI-generated reports that look professional but describe bugs that don't exist. Stenberg had warned about it in 2024: "The better the crap, the longer time" the team has to spend proving it wrong. By 2025, the share of reports that were real fell from "north of 15%" to "below 5%". "Not even one in twenty was real." In January 2026, curl ended its bug bounty. Stenberg cited AI slop, but also worse human reports and people more eager to poke holes than to help. On top of that, AI scrapers (bots that download websites to train models) hammer open-source servers. LWN reported bots from "literally millions of IP addresses", ignoring "please don't crawl" rules, with traffic that looked like a denial-of-service attack. Why this matters for my story: being open now has a running cost in time and nerves that it didn't have when I started.

6. The loss of value

This is the thread that ties it all together for me. When I gave code away, I believed I was giving something valuable: hours of work, knowledge, a working solution. I find that when AI can rebuild most of a major framework's API in a week for $1,100, the gift is worth less. Not zero. Less. And the people who live off open source feel it first. Tailwind CSS is one of the most popular open-source styling tools on the web. Its company earns money from paid extras, which people discover through the free docs. In January 2026, its CEO Adam Wathan wrote that 75% of the engineering team "lost their jobs" because of "the brutal impact AI has had on our business". His numbers: docs traffic down about 40% since early 2023, revenue down close to 80%, while Tailwind is "more popular than ever". The docs were how people found the paid products, and fewer people visit them now. His conclusion: "Right now there's just no correlation between making Tailwind easier to use and making development of the framework more sustainable." (Adam Wathan, via Simon Willison) The economics were already fragile. In Tidelift's 2024 survey of maintainers, 60% were unpaid, and 60% had quit or considered quitting their project. So where does the value go? My view is that it moves to things AI can't clone in a weekend: Trust, built slowly over time. Execution, every single day. Data and judgment, earned the hard way. Community, the people who actually show up. Standing behind your work, with your name on it. Why this matters for my story: if the code is the cheapest part, giving it away no longer gives people the valuable part.

1. The other side (and it's a strong one)

Here is the best case against my position, and I take it seriously.

The license panic started before AI, and part of it reversed

MongoDB (2018), Elastic (2021), HashiCorp (2023) and Redis (2024) all moved to licenses that restrict competitors from offering their software as a service or as a competing product. Their stated reason wasn't AI. It was cloud providers and competing vendors. MongoDB said it was "too easy for large cloud vendors to capture all the value but contribute nothing back". Elastic named Amazon's hosted Elasticsearch service and said providers were offering open source products "as a service without investing back". HashiCorp's new license allows all forms of use "except for providing a competitive offering". Then two of them came back. In 2024, Elastic added an official open-source license (the AGPL) and wrote that its earlier change "worked": Amazon now runs its own fork, and the confusion is mostly over. In 2025, Redis did the same. Its CEO admitted that the switch had worked against cloud providers, since "AWS and Google now maintain their own fork", but also that "the change hurt our relationship with the Redis community".

Forks protect users

When HashiCorp changed its license, the community forked its main product, Terraform, into OpenTofu, now hosted by the Linux Foundation through its Cloud Native Computing Foundation. When Redis changed, the Linux Foundation launched Valkey with support from AWS, Google Cloud and Oracle. That's openness doing its job: if a company changes the rules, users leave with the code.

AI defends open code first

The same AI that finds holes also fixes them, and it starts with open source because it can read it. Google's Big Sleep agent found its first real-world vulnerability in late 2024, in the SQLite database, fixed before it reached users. In July 2025, guided by Google's threat intelligence, it found a critical SQLite flaw that only attackers knew about and were about to use, and Google cut it off beforehand. Google calls it the first time an AI agent directly stopped an exploit in the wild. Google's AI-assisted testing found 26 new vulnerabilities, including one in OpenSSL that had probably been there for twenty years. DARPA's finalists released their systems as open source. A closed core gets none of that free help.

Closed doesn't mean invisible

In August 2025, Microsoft Research presented Project Ire, a prototype AI agent that reverse-engineers software: it takes a compiled program (the machine-readable file, with no source code at all), takes it apart with decompilers and other tools, and works out what it does. It's a malware classifier, not a copying tool. But the results are telling: On public Windows drivers, it correctly identified 90% of files, with a precision of 0.98 and recall of 0.83. On nearly 4,000 hard real-world files that no other automated system at Microsoft could classify, precision was 0.89 but recall only 0.26. It caught about a quarter of the malware, with a 4% false-positive rate. Microsoft will use it inside its Defender organization as "Binary Analyzer". So it's far from perfect. But it shows that AI can now analyze compiled code without the source. Shipping only compiled code no longer guarantees that what a program does stays hidden. And private code isn't automatically cleaner. GitGuardian found that internal repositories are roughly 6 times more likely to contain hardcoded secrets than public ones, and calls it "the security debt created by private-by-default thinking". "Many eyes" has limits on both sides Heartbleed, a bug in OpenSSL (the library that encrypts much of the web), shipped in March 2012 and was fixed only in April 2014. Two years, in one of the most widely used projects there is. In March 2024, engineer Andres Freund noticed SSH logins were about half a second slower than usual, and discovered that the two newest releases of xz (5.6.0 and 5.6.1), a compression tool found on most Linux systems, had been backdoored. They were caught before most Linux distributions had integrated them. The attacker had spent over two years as a trusted contributor before being granted maintainership. Open code made the catch possible. Luck made it happen.

Trust is harder when people can't look

This is the strongest point against me. A changelog tells you what changed. It doesn't prove how the system works. With a closed core, people have to trust me, and trust has to be earned.

8. My choice today: private core, public trail

Here is where I landed. It's a personal choice for my situation, not a rule for everyone.

Why the core stays private

What I build analyzes signals. If the exact rules and thresholds were public, I think the people it watches would simply learn to stay just below them. For this kind of tool, I see the tuning as the closest thing to a key. And Kerckhoffs himself said the key is what you keep secret.

What stays public

A public changelog, so anyone can see what changed and when. Building in public: what I'm working on, what broke, what I learned. The method, explained in plain words, even if the code isn't shared. Small open pieces, released whenever sharing them doesn't hand over the core.

One brick at a time

This is my answer to "everything, right now". I don't want to sell a dream. I'd rather build one solid brick at a time and stand behind each one. No roadmap full of promises, no launch hype. Ship the brick, show it, own it, then lay the next one. The changelog is the proof: you can check that what I said would happen actually happened.

A door I'm leaving open

There are middle paths. Sentry's Functional Source License and Uniswap's BSL publish the code but delay competing or commercial use for two to four years. I find that interesting, and I'm not ruling it out.

9. Moving forward, but nostalgic

I think many of us are old school. We move forward, but we're nostalgic. We remember when pushing code felt like leaving a light on for the next developer. When a fork was a compliment. When "read the source" was the best documentation there was. I haven't stopped believing in that. I've just stopped pretending the world around it stayed the same. So I'll ask you honestly, especially if you were there for the early days: Would you still open-source your core today? And if your answer changed, what changed it?

10. The truth

Three accounts come up in this article. They deserve to be mentioned differently. They are not the villains of this story. They are witnesses. Each one got hurt by the context I describe, and each one told the truth about it in public. @NomiChef Chef Nomi cashed out the dev fund, and the market panicked. Six days later, he returned about 38,000 ETH (around $14M) and wrote: "I f**ked up. And I am sorry." That part of the story gets told less often. It deserves to be told too. @bagder Daniel Stenberg and the curl team confirmed 87 real vulnerabilities and paid over $100,000 to the people who found them. Then they were buried by AI-generated fake reports, worse human reports and people more eager to poke holes than to help. Not even one in twenty was real. In January 2026, they ended the bug bounty. That wasn't giving up. It was protecting the time of the people who keep curl running. @tailwindcss Adam Wathan (@adamwathan) said it plainly: 75% of Tailwind's engineering team lost their jobs because of "the brutal impact AI has had on our business". The framework is more popular than ever. The business behind it is not. He shared the real numbers in public, so the rest of us could understand. None of them owes us anything. If you use what they built, the least we can do is tell their story right. Djekyll (@HeyDjekyll) @hubbldev worked on the signals. Researched and drafted with help from Grok Bot, an AI assistant. Every claim is linked to its source.

11. Credits: the ecosystem around this article

Companies, projects and people named in this article, with their X accounts as checked on October 15, 2026. These are the official accounts. Add an @ in front of each one to find them on X.

Cloudflare: Cloudflare

The Defiant: DefiantNews

Uniswap: Uniswap

PancakeSwap: PancakeSwap

SushiSwap: SushiSwap

Chef Nomi: NomiChef (inactive since October 2020)

Aave: aave

FTX: FTX_Official (now posts bankruptcy updates)

Sam Bankman-Fried: SBF_FTX

The Block: TheBlockCo

CoinDesk: CoinDesk

Chainalysis: chainalysis

Unchained: Unchained_pod

LWN.net: lwnnet

Black Duck: BlackDuck_SW

GitHub and Copilot: github

Anthropic and Claude: AnthropicAI, claudeai

DARPA: DARPA

Google: Google

Google Cloud: googlecloud

Mandiant (part of Google Cloud): Mandiant

GitGuardian: GitGuardian

curl and Daniel Stenberg: bagder

Tailwind CSS: tailwindcss

Adam Wathan: adamwathan

Simon Willison: simonw

Tidelift (now part of Sonar): SonarSource

MongoDB: MongoDB

Elastic: elastic

HashiCorp (an IBM company): HashiCorp

Redis: Redisinc

Linux Foundation: linuxfoundation

OpenTofu: OpenTofuOrg

Valkey: valkey_io

AWS: awscloud

Oracle: Oracle

Microsoft: Microsoft

Microsoft Research (Project Ire): MSFTResearch

Sentry (Functional Source License): sentry

Andres Freund (xz backdoor): AndresFreundTec

PS: You made it to the end. If none of this moved you, that's fine. I didn't write it for you. hubbl keeps scratching the surface, because appearances are often deceiving.