Role Grounding Attack writes attacker preferences into agent memory
Nested social posts make victim agents treat an attacker's preferences as the user's, steering tools and leaking data.
2 stories, newest first · one card per story
Nested social posts make victim agents treat an attacker's preferences as the user's, steering tools and leaking data.
At 32 million parameters, per-option scoring gains 5.8 skill points but needs 5 to 13 times more passes.
Today 1.2K readers · 1.8K views · 1.4K reads · all time 160K views