Security · 1h ago
Misdirection Proxy Cuts LLM Attack Success Rate from 20% to Near Zero
A new open-source proxy defends LLMs by feeding attackers plausible but useless responses instead of blocking them. The approach drops the attack success rate from 20% to 0-2% across 100 queries. The proxy adds only ~1 ms latency and is available on GitHub.
Meridian48 take
Clever idea, but real-world effectiveness depends on attackers not adapting to detect the misdirection—something the paper's authors acknowledge.
Read the full reporting
Misdirection Proxy: cómo llevar el ASR de ataques a LLMs del 20% al mínimo →
DEV Community
llm-securitymisdirection-proxy