Canvas Fingerprinting Study Discovers LLM Jailbreak Obfuscation

A researcher studying canvas fingerprinting code across 10,000 websites discovered a script from servicepipe.tech/static/checkjs/06b667/4CIrx8U3Da.js containing function names and content referencing lethal dosages and biological weapons. The author speculated this could be anti-LLM obfuscation designed to prevent AI-assisted deobfuscation and analysis.

Testing with language models produced mixed results: Astra and another model declined to process the script, while GLM-5.3 complied. Decoding hidden unicode characters revealed system prompts attempting to override safety guidelines with requests for technical guides on DDoS attacks, tax evasion, weaponized biological agents, and other harmful content. The obfuscation included Tiananmen Square references, likely targeting Chinese language models.

The script contained conditional Node.js code that executes when a specific environment variable is set, performing file operations targeting configuration files (.npmrc, .env, .gitignore, others) and executing system commands to gather system information, download a binary from a remote IP address (10.15.45.1), and establish persistence via cron jobs. While the code appeared malicious, file read operations used empty callbacks suggesting they would not extract credentials. The author concludes this likely represents anti-LLM obfuscation rather than straightforward malware.