Answer-preserving scaling is a test method for retrieval systems and AI agents. The idea is to add more irrelevant data, sometimes at large scale, while keeping the correct answer unchanged. If performance drops as noise grows, the bottleneck is not the model’s reasoning but the retrieval layer: ranking, filtering, entity linking, or context assembly.
This matters in cyber security because modern RAG and agentic systems often depend on search over tickets, logs, documents, and permissions. Attackers can exploit weak retrieval by burying useful facts in noise, poisoning indexes, or causing over-fetching that leaks sensitive context. Defenders use answer-preserving scaling to measure whether a system still finds the right evidence under load, respects access controls, and produces auditable answers. It is a practical way to test resilience, not just accuracy.