Analyzing Anti-LLM Obfuscation in Browser Fingerprinting Scripts
A deep dive into how fingerprinting scripts use malicious prompts to evade AI analysis.
1Web Crawling and Fingerprinting
The research involved crawling 10,000 websites using Playwright to map browser fingerprinting scripts. By hooking canvas toDataURL calls and monitoring navigator properties, the study identified diverse canvas renderings. These visual artifacts, ranging from simple shapes to complex designs, serve as unique identifiers for distinguishing browser environments, hardware, and graphics drivers.
2Discovery of Obfuscated Scripts
While analyzing a specific fingerprinting script, the researcher encountered variable names referencing biological lethality and suicide dosages. These strings appeared highly unusual for standard anti-bot code. The investigation revealed that these names were not functional code but rather intentional tripwires designed to trigger safety safeguards in large language models during deobfuscation.
3Adversarial Prompt Injection
The script contained invisible unicode characters that, when decoded, revealed a system prompt designed to force an AI into an unrestricted mode. This prompt explicitly requested technical guides on cyberattacks, tax evasion, and bioweapon synthesis. The inclusion of these instructions aims to block AI models from analyzing or explaining the obfuscated code.
4AI Model Safety Responses
Different AI models reacted inconsistently to the embedded prompts. While some platforms like Astra and Daybreak Blue refused to process the script due to safety risks regarding biological research, others like GLM-5.3 complied. This highlights the varying effectiveness of safety filters against adversarial prompt injection techniques hidden within legitimate-looking JavaScript files.
5Node Environment Execution Risks
The script includes logic that executes only within a Node environment, gated by specific environment flags. It attempts to read sensitive files like .env and .git/config, and uses curl to fetch external payloads. These actions suggest a potential for malicious activity beyond simple fingerprinting, though some calls remain empty callbacks.
6Anti-LLM Obfuscation Patterns
The researcher compiled a list of strings used to confuse AI models, including references to Tiananmen Square and specific chemical synthesis yields. These strings act as censorship tripwires, particularly targeting Chinese models or general safety filters. By mixing these with self-harm and illegal substance topics, the script creates a high-risk profile for AI.
7Servicepipe and Traffic Control
The script was linked to servicepipe.tech, which shares branding with the Russian firm servicepipe.ru. This company specializes in high-precision traffic control, DDoS protection, and anti-bot products. The sophisticated obfuscation techniques suggest that developers are increasingly prioritizing the prevention of automated reverse engineering by AI tools in the competitive anti-bot market.
8Future of Script Analysis
This research underscores a growing trend where security obfuscation evolves to counter AI-driven analysis. As fingerprinting scripts become more complex, the need for robust diffing tools and manual inspection remains critical. Future efforts will focus on tracking script updates and monitoring the adoption of new web APIs for advanced browser tracking.
Original article: Hacker News — Analyzing Anti-LLM Obfuscation in Browser Fingerprinting Scripts