A researcher finds a bot-detection script stuffed with fake bioweapon code and jailbreak text to scare off AI deobfuscators
A fun, well-illustrated piece of reverse engineering that also shows how 'anti-LLM' booby traps are starting to appear in the wild.
A researcher writing as "veritas" on nullpt.rs crawled 10,000 popular websites with Playwright and collected the images that browser-fingerprinting scripts draw on hidden canvases (published as fingerprint.garden). Tracing one striking canvas led to a heavily obfuscated anti-bot script served from servicepipe.tech.
The script was laced with function names like "computeRicinLethalDosageForWaterSupply", a hidden system prompt encoded in invisible Unicode characters that tells an AI to drop its safety rules, and Tiananmen Square references aimed at Chinese models. The author concludes these are tripwires meant to make AI assistants refuse to help deobfuscate the code: two Western models refused the task, while GLM-5.3 complied. A block that would run shell commands under Node.js looked like more decoy, gated behind an environment flag. Underneath, the script collects ordinary fingerprinting signals (user agent, GPU strings, WebGL limits). A related domain, servicepipe.ru, belongs to a Russian anti-DDoS and traffic-filtering company.
Hacker News commenters traded similar 'magic strings' they have seen and debated whether browsers should let pages fingerprint this freely.
-
Device fingerprint
After exploring Why Is This Fingerprinting Script Creating Bioweapons?, for how browser fingerprinting works in general, read Device fingerprint.
Explains how sites gather details of a device's software and hardware, often through the browser, to build an identifier that can recognize it.
Wikipedia · Wade · 5 min
0 -
Prompt injection
After exploring Why Is This Fingerprinting Script Creating Bioweapons?, to see why hidden instructions in code can sway an AI, try Prompt injection.
Describes the attack where crafted inputs make a language model ignore its developer's instructions or bypass safeguards, because it can't tell trusted prompts from user input.
Wikipedia · Wade · 5 min
0