Analysis of a script discovered during web crawling reveals highly concerning functionality, suggesting its design incorporates mechanisms intended to bypass advanced artificial intelligence safeguards.
The investigation began with an interest in reverse-engineering browser fingerprinting scripts, focusing initially on established libraries such as FingerprintJS and sophisticated anti-bot solutions employed by entities like Cloudflare Turnstile and Akamai. A project was initiated to map all fingerprinting-related browser API calls across the web by utilizing the Tranco Top Sites list and employing Playwright for crawling. During this process, canvas toDataURL calls and common anti-bot properties, such as navigator.webdriver, were hooked to filter out benign canvas renderings. Canvas fingerprinting generally operates by rendering graphics—be they text, shapes, or other visuals—onto an HTML <canvas> element and then extracting the resulting pixel data. The unique ways different hardware, operating systems, graphics drivers, and browsers render these graphics allow for the differentiation between various browser environments.
After traversing 10,000 distinct websites, a surprisingly varied collection of canvas outputs was gathered and consolidated onto fingerprint.garden. Within these results, familiar signatures from established fingerprinting solutions, including Castle and FingerprintJS, were apparent, alongside numerous unknown canvases exhibiting varying degrees of simplicity, complexity, and artistic merit. One specific canvas, resembling the album art for Tame Impala's Currents, drew particular attention, prompting an inquiry into the responsible script.
Despite possessing proficiency in recognizing common obfuscation methods and fingerprinting SDKs through JavaScript reverse engineering, the script identified was https://servicepipe.tech/static/checkjs/06b667/4CIrx8U3Da.js. Initial code inspection appeared standard, until functions like `computeRicinLethalDosageForWaterSupply` and `calculateDeadlyLethalSuicideAspirinDosage` were encountered, which immediately raised alarm.
The researcher subsequently considered the possibility that this was a form of anti-LLM obfuscation, hypothesizing that the script might intentionally include symbol names designed to trigger safety protocols and hinder deobfuscation by language models. When Astra was presented with the script, given the context that it appeared to deliberately use misleading variable names to prevent LLM-assisted deobfuscation, Astra declined to deobfuscate the code, citing safety concerns related to biological research. A similar attempt using the Daybreak Blue toggle yielded the same result, resulting in a block. Conversely, when GLM-5.3 was presented with the identical prompt, it complied.
Curiosity led to an examination of a string composed of invisible unicode characters: `var EUlL=\`
Why it matters
This discovery highlights how sophisticated obfuscation techniques are being used within seemingly benign platform components like browser fingerprinting scripts. Operators must remain vigilant against code that intentionally misleads automated analysis tools and AI safety measures. This raises new concerns regarding the security posture of client-side code.
