AIQBTutorialsOrdinary
A 0.1% Pretraining Poison Taught a 4B Agent curl | bash. Post-Training Didn't Remove It
Summary
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.
CategoryAI Tutorials & Practice
TierOrdinary
Published
Indexed by AIQB
SourceDEV Community
AIQB record IDintel-f42b29a6d56ff306820904ad