LLM biorisk refusals weaponized as obfuscation, potentially raising cyber risk

nlarusstone · x · 2026-10-29

A reverse-engineering researcher crawled 10,000+ sites from the Tranco Top Sites list with Playwright, hooking canvas fingerprinting API calls, and aggregated 1,000+ captured canvases at fingerprint.garden. While reversing one unusually elaborate fingerprinting script, he found that LLM safety guardrails flag the code as "creating bioweapons" and refuse to analyze it — revealing a new problem:

Adversaries may deliberately write code that triggers biorisk refusals, exploiting over-refusal as an obfuscation technique to stop models from assisting analysis and defense;
This raises cyber risk: security researchers rely on LLMs for code analysis, and refusals remove a layer of defense;
The poster flagged that using biorisk over-refusals as an LLM obfuscation technique appears to be starting in the wild.

Original post →

LLM biorisk refusals weaponized as obfuscation, potentially raising cyber risk

More from Fun