INFUSE tests 19 language models on 56 prompt injection attacks
The mean attack success rate is 19.6% when a language model examines the result.
Claimed, not confirmed
Researchers made INFUSE, a benchmark for indirect prompt injection in long agent tasks. It has 40 tasks and 56 attacks. In tests on 19 models, the mean attack success rate is 19.6%. A prompt-level defense cuts the mean rate from 58.2% to 14.3%.
Sources
Posted
