Poisoned function names trick AI models into refusing code analysis, bypassing traditional obfuscation
paul_cal · x · 2026-10-31
Security researcher veritas found JavaScript stuffed with bioweapon-themed function names like computeRicinLethalDosageForWaterSupply — while the code merely collects browser fingerprints. Some models refuse to analyze it on safety grounds; GLM-5.3 succeeded. This new obfuscation targets LLM safety mechanisms rather than compilers, and stricter safety alignment may make models more susceptible.
Related event: Anti-Bot Scripts Weaponize AI Safety Filters with Bioweapon Function Names(3 posts)→
More from Fun
- The AI 'stages of grief': timelines split over whether AI can do your job — IgorCarron · 2026-10-31
- Anthropic Design Shows Off Months of Small Claude App Polish and Dad Jokes — dfeinition · 2026-10-31
- Startup loses $4K overnight as bot farm drains free AI credits to make a short drama — R4nsack · 2026-10-31
- Casimir effect diagram vs. Rothko's Black on Maroon: two kinds of inscrutable depth — jjvincent · 2026-10-31
- The internet stopped being fun because everyone lives online now, leaving nothing to post — Kyrannio · 2026-10-31
- Gemini happily lists piracy sites to 'avoid' when a Redditor asks — gameplaydex · 2026-10-31