• Sources: primary, discussion
  • Summary: A researcher crawled 10,000 sites from the Tranco list and traced one canvas fingerprint to a script on servicepipe.tech. The script uses function names such as computeRicinLethalDosageForWaterSupply, a jailbreak prompt hidden in invisible Unicode tag characters, and Tiananmen references, and one frontier model refused to deobfuscate it while GLM-5.3 complied. Underneath, it collects ordinary navigator and WebGL signals, and a Node-only block that looks like credential theft appears to be decoy code.
  • Why it matters: Shipped JavaScript can now deliberately trip model safety filters, so teams that use LLMs for deobfuscation or code review should expect refusals and keep non-LLM fallbacks.