Concept

Claudini

A white-box adaptive attack method against LLM prompt injection that uses gradient information to optimize inputs in order to generate adversarial injections.

Timeline (last 90 days)

Oct 28

Against the white-box attack Claudini, attackers need to consume about 3 times the FLOPs to reach the same ASR on HoS as on the baselines.

Unverified50%

All Facts (1)

Unverified

Against the white-box attack Claudini, attackers need to consume about 3 times the FLOPs to reach the same ASR on HoS as on the baselines.

50%

Source Articles

Attention Gating: A New Defense — How HoS Makes Prompt Injection Attacks FailOct 28