TL;DR: When attacks on search agents evolve to adapt to the defenses, they cut accuracy far more than fixed attacks. Static benchmarks may be underestimating the risk.
The Web Doesn't Sit Still: Adversarial Self-Evolving Attacks on Search Agents
Key points
・🧪 Evaluates RAG poisoning with evolving attacks instead of fixed templates
・🧬 CREO: contrasts successful and failed reasoning traces, and an LLM crosses over and mutates strategies
・🔁 A bi-level setup where the defender discloses last round's attack patterns in its prompt
・📉 GLM-5.1 on MuSiQue: base Acc 82 drops to 41 under AuthChain and 16 under CREO
・⚠️ Mean TASR is 60.5–74.6% for CREO versus about 20% for FactSwap
・😵 A new metric, APR, also measures "cognitive paralysis", where agents refuse to answer after confusing contradictions
・🛡️ Keyword and perplexity filters barely help; even the effective defense leaves large residual risk
The evaluation covers multi-hop QA only, and the defender knows the strategies. Still, it suggests search-agent safety evaluation should assume adaptive attackers.
#
AISecurity# #
SearchAgents#