SEAL cuts AgentDojo prompt-injection success from 34.6% to 0.1%
Joint task completion without attack success rises from 51.4% to 77.5%, using about 127 times fewer defense tokens than PISanitizer.
52
Joint task completion without attack success rises from 51.4% to 77.5%, using about 127 times fewer defense tokens than PISanitizer.
Today 468 readers · 529 views · 253 reads · all time 168K views