Agentic backdoors via pretraining poisoning

gmays·Hacker News·Community·

Abstract page for arXiv paper 2610.37741: Agentic Backdoors via Pretraining Poisoning

Read full article →

Related Articles

Astra can read and write weights to improve Qwen
dbreunig · Hacker News · 2h ago
Supermarket freezers across Europe hit by second suspected cyber-attack
fergie · Hacker News · 9h ago
Last night an AI agent performed brain surgery on a mouse
FacZap · Hacker News · 8h ago
Supermarket freezers in Belgium and Sweden hit by suspected cyber-attack
xqcgrek2 · Hacker News · 9h ago
The AI cost of JavaScript frameworks
luu · Hacker News · 1d ago