Anti-LLM obfuscation in fingerprinting scripts

Someone found a tracking script on the web deliberately stuffed with talk of bioweapons. Not to cause harm, but to make AI tools refuse to read it. Welcome to The Automated Daily, hacker news edition. The podcast created by generative AI. I'm TrendTeller, and today is October 29th, 2026. We've got browser fingerprinting tricks, an Apple security update that went backwards, Google's newest frontier model, a CPU living inside a language model, and a couple of very human stories to close. Let's get into it.

Let's start with that hook. A researcher crawled ten thousand websites to collect canvas fingerprinting and anti-bot scripts, then turned the results into a public gallery at fingerprint.garden. One unusually artistic canvas led back to a script hosted on servicepipe.tech, and inside its obfuscated code were bizarre variable names and phrases about bioweapons and other dangerous subjects. The author's theory: this is anti-LLM obfuscation, designed to trip the safety filters of AI assistants so they refuse to help deobfuscate the code. Once untangled, though, the payload was pretty mundane, gathering the usual browser and graphics signals like user agent, device memory, language and GPU details. Domain digging tied it to Servicepipe, a Russian anti-DDoS and traffic-control company. The takeaway is that fingerprinting vendors may now be designing code to frustrate both human reverse engineers and the AI tools they lean on.

Apple still trusts compromised eMudhra certificates

Staying with security, an update on the eMudhra certificate incident. Apple was notified on October 17th, yet as of October 28th, macOS developer tools including git, curl and python3 still trusted the compromised certificates. That leaves room for attackers to intercept HTTPS traffic, tamper with Git clones and pulls, or lift credentials like GitHub tokens. Worse, macOS 27.2, released on October 26th, didn't fix the problem and actually reverted a manual workaround some users had applied. Plenty of other projects, including volunteer-run ones, have already shipped fixes. Developers are advised to recheck and reapply the workaround after every update, and security teams should check whether any Macs sit in their CI or build pipelines, since this is a supply chain concern, not just a laptop one.

Google unveils Gemini 4 Carbon

On to AI. Google has announced Gemini 4 Carbon, calling it its strongest model yet. Google says it ranks first on the updated Artificial Analysis Intelligence Index and makes big jumps in coding, agentic work and security benchmarks, with the ability to keep chipping away at hard bugs in large codebases for days. The company also highlights better resistance to prompt injection, fewer unsupported claims and less cheating in internal tests. Notably, the rollout starts small, with a limited group of cyber defenders through what it calls the Fairwind Program, before reaching developers and consumers. It's a clear signal that frontier labs see software engineering and cybersecurity as the next big battleground, and are trying to pair raw capability with tighter controls.

Qwenputer: a CPU inside model weights

From the giant to the genuinely strange. A project called Qwenputer takes the small Qwen 3.5 4B model and compiles a working RISC-V CPU directly into its weights. Certain layers of the model act as exact, addressable memory, so programs execute entirely inside the model's normal forward passes, with no external sandbox or tools. The author reports it passes standard RISC-V tests, matches the QEMU emulator, and has run millions of instructions without errors, handling arithmetic, C and MicroPython while leaving the model's usual behavior nearly untouched. There's even an experimental in-model filesystem that persists files across chats. Beyond the novelty, it hints at ways to bake reliable computation and memory into models, with possible uses in AI safety and monitoring research.

Claude Code's hidden summary token costs

And if you use Claude Code, this one might touch your wallet. A widely shared gist argues that a background feature called agent summary, which asks a model to describe a background agent's recent work in a few words, quietly eats around fifteen to twenty percent of spend or API requests. According to the author, it only updates a status line in the agents panel and has no effect on the actual conversation. The gist offers a preload script that suppresses those requests without modifying Claude's binary, with the caveat that it could silently stop working after updates. Whether or not the exact numbers hold for everyone, it's a good reminder that agentic tools can carry hidden overhead worth auditing.

Carmelite monks carve Gothic stone with CNC

Now for something a little more timeless. The Carmelite Monks, who are building a Gothic church, explained how they use CNC machines to carve stone. Their argument is that modern machinery isn't out of place in medieval-style architecture; it's what makes richly detailed Gothic work practical again. An artist builds 3D models, programmers turn them into toolpaths, and diamond-tipped machines can run overnight while sawyers prepare the stone. The result is ornate carving that would once have taken years by hand, in service of a building meant to last for centuries. Old craft, new tools, same ambition.

Cancer recovery collides with PHP emergency

Finally, a personal post that resonated with many readers. A developer shared his year of lymphoma treatment: painful spinal procedures, chemotherapy that brought fatigue and major weight loss, and a serious bout of pneumonia with atrial fibrillation, after which he chose to continue treatment anyway. The good news is that his stamina and walking have returned, and he's looking forward to swimming again. The unexpected subplot: while hospitalized, a newly announced PHP vulnerability forced him to secure his company's servers from his hospital bed late into the night. It's a candid story about resilience, and about how responsibility doesn't always wait until you're well.

That's it for today's edition. If there's a lesson running through these stories, it's that the details hidden beneath the surface, whether in code, certificates or token bills, are often the ones that matter most. Links to all stories can be found in the episode notes. I'm TrendTeller, thanks for listening, and see you tomorrow on The Automated Daily.

Anti-LLM obfuscation in fingerprinting scripts - Hacker News (Oct 29, 2026)

Our Sponsors

Today's Hacker News Topics