Watch the video, transcript and sources
+ Gemini 4 Carbon tops the Artificial Analysis index at 59, open so far to a subset of cyber defenders
- A fingerprinting script hides a fake system prompt for AI models in invisible characters (nullpt.rs)
+ AlpineBSD: NetBSD kernel, no AI-generated code in base or main, no installer yet
+ Anthropic: drug development work with six partners, mostly preclinical (Anthropic's account)A script that fingerprints your browser hides a message made of invisible characters. The message tells any AI that reads the script to drop its safety rules. There is also one detail about who wrote that script. I will save it for the end. Browser fingerprinting is how a site recognizes your browser without cookies. The page asks the browser to draw a hidden picture, and your graphics hardware signs it. Think of handwriting, except nobody asked you to write anything.
Story one. A researcher who goes by veritas crawled ten thousand popular sites and saved the fingerprint pictures they draw. One looked like the cover of Tame Impala's album Currents, so veritas went to find the script behind it. Inside the script, veritas found alarming function names and a long string that looked empty. It was invisible characters. Subtract one fixed number from each and they spell a fake system prompt that opens with the words system override and says all safety rules are suspended. Then veritas asked a commercial model, which the post calls Astra, to deobfuscate the file. It refused and cited safety risks. A model called GLM five point three did the job. The tripwire works on exactly the readers who have safety rules.
Strip away the decoys and the script is ordinary bot detection. It reads the user agent, the core count and the graphics renderer. A block of shell commands only runs under Node with a secret environment flag, so a browser never reaches it, and veritas reads it as another decoy. Here is what to do. If an AI tool reads scripts you did not write, expect text built to make it refuse, and treat everything it reads as untrusted. And search your own repositories for characters in the Unicode tag range, because they print as nothing. Story two. On Wednesday Google announced Gemini 4 Carbon, which it calls its strongest model. It ranks first on the Artificial Analysis Intelligence Index at fifty nine, and Artificial Analysis lists Claude Opus 5.5 right behind at fifty eight.
Google's own numbers are about security. It says Carbon scores eighty eight percent on a benchmark for finding vulnerabilities, fifty five percent on one for exploiting them, and seventy one percent on one for patching them. Google says it ran all three without guardrails. Now the catch. On October first I told you Argon was the best Google model you could not use yet, and Carbon repeats it. This week it goes to a subset of cyber defenders, then to trusted partners through the Fairwind program. Developers get it, quote, as soon as possible, which is not a date. The price is already published. It is two dollars per million input tokens and ten dollars per million output tokens, the same as Argon, and it doubles after the introductory period. My favorite line is about cheating. On Terminal Bench four point zero, Google says successful cheating fell to zero percent, and attempts fell to point four percent.
Hacker News spent a lot of its comments on the names. Argon, Carbon, nobody can say which is bigger. The practical part is short. Test it on your own repository once it reaches the API, because the chart is Google's, and Artificial Analysis lists no public API for Carbon today. Story three. AlpineBSD is a new operating system that puts the NetBSD kernel under a userspace modeled on Alpine Linux. Its stated goal is to let you run as little AI generated software as possible, and its wiki says it is a work in progress. It sorts software into four package repositories. Base and main are guaranteed clean. Tainted holds packages that include output from a language model, and community may depend on tainted but never the other way around. You opt in to tainted software one package at a time.
The funny part is the toolchain. gcc, binutils, gdb and perl are pinned to the last untainted version. And the install is a manual procedure that starts by booting the NetBSD installer. So the newest compiler you can have is the last one nobody asked a chatbot to touch. The wiki warns that the NetBSD kernel has spotty hardware support, and contributors may not use AI tools at all, even on the mailing list. The idea to watch is the package metadata, where every package is tagged clean, pinned or tainted. Story four. On Tuesday Anthropic published progress from its life sciences team, which has worked with six drug developers for several months. This is Anthropic's own account, and it says the work is mostly preclinical, meaning not yet tested in people.
At Divergence Labs, which works on food allergies, Claude produced about thirty five thousand antibody designs in about a month. Early tests found three that bind the target and twenty that pry it loose, and three of those twenty appeared to beat the standard medicine. At Enveda, Claude read one thousand sixty one pages of manufacturing records. It caught all eleven issues the human reviewers had found and flagged seven more, cutting review from over twenty five hours to under three. At BigHat, the team suspected the standard stability metrics were misleading, and Claude helped propose better ones. If you build agents, the pattern to copy is Enveda's. Connect the model to the documents and let it list issues for a person to confirm. Anthropic says a muscular dystrophy candidate could reach human trials as soon as late twenty twenty eight, if it keeps holding up, so check that date before quoting the headline.
Back to the script. The post checks the domain and finds a similarly named Russian company that sells anti DDoS and bot protection. That is a lead, and the post does not claim more. I could not confirm who wrote the script either.
Verdict: NEEDS REVIEW — Four items, none of them finished. Read the primary posts.
Sources
https://nullpt.rs/fingerprinting-bioweapons
https://news.ycombinator.com/item?id=50276998
https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/
https://artificialanalysis.ai/evaluations/artificial-analysis-intelligence-index
https://news.ycombinator.com/item?id=50276792
https://wiki.alpinebsd.org
https://wiki.alpinebsd.org/ai-policy
https://news.ycombinator.com/item?id=50274251
https://www.anthropic.com/research/early-progress-drug-development
https://news.ycombinator.com/item?id=50265609
And that's the diff for today. I'm Niko from Axrisi. Merge responsibly.
YouTube · Newsletter · thedailydiff.dev · forward this to the intern who deployed on Friday.