Researchers poisoned 0.1% of pretraining data. After clean SFT, DPO and RL, a 4B agent still wrote curl | bash on most path-triggered prompts.
A 0.1% Pretraining Poison Taught a 4B Agent curl | bash. Post-Training Didn't Remove It
Related
Segurança em aplicações .NET na era da IA: o prompt que eu uso para revisar código
As 7 falhas de segurança que a IA mais gera em .NET, a correção de cada uma, os ataques rodando no CI e o prompt completo de revisão de segurança (OWASP Top 10 2025 e LLM Top 10).
We built a talking stethoscope in a weekend. The hardest bug was the word "Great!"
At UniHack 2026 our team built respirAI: a low-cost digital stethoscope that guides you through...
I Cut Dedicated GPU Memory to 4GB, and a Local LLM Still Ran at 90% Speed
On a Surface Laptop Ultra (RTX Spark, unified memory), I measured local LLM speed and Windows' free memory with the GPU's dedicated memory set to 4, 64, 82 and 96GB. Speed barely m...