Anti-bot scripts hide ricin-dose function names to trip AI safety filters
matthew_d_green · x · 2026-10-29
A browser-fingerprint researcher who crawled 10,000 sites reverse-engineered anti-bot scripts and found obfuscation aimed at AI, not humans: function names deliberately reference ricin lethal doses, dirty-bomb blast radii and fentanyl precursor yields to trigger safety refusals when devs paste code into an LLM. GPT-6 Astra indeed refused; the script also hid an invisible-character jailbreak prompt. Testing GLM-5.3, it happily deobfuscated everything anyway.
More from Fun
- Nvidia interview question goes around: hand-trace the Z-algorithm on a string — cneuralnetwork · 2026-10-29
- A 20-line jailbreak prompt tried to hijack open-source Qwen — and the model saw right through it — huangyun_122 · 2026-10-29
- AI confidently gives the wrong answer in viral fail clip — thermologic_ · 2026-10-29
- Alexandr Wang responds to viral meme with 'I thought we were all friends' — minchoi · 2026-10-29
- Deep learning optimizer papers: theoretically flawless, illustrated with Morse code — cloneofsimo · 2026-10-29
- Runway CEO revisits 2019 AI short film, making 'Will Smith spaghetti' look HD — c_valenzuelab · 2026-10-29