BRANCH bypassed 6 AI guardrail systems in all 120 tests
The authors write that its attacks also succeeded on 29 guardrails that it had not seen.
Claimed, not confirmed
Researchers propose BRANCH, a method to bypass guardrail systems that use many scanners to find malicious instructions. The authors write that it gave a 100% attack success rate on 6 guardrail systems in 120 tests. The number of queries was 72% less and the time was 4.5 times less than for known methods. The attacks also succeeded on 29 guardrails that it had not seen, 8 of them commercial black-box guardrails.
Sources
Posted
