For over 14 years, Backblaze has been collecting the failure rates of hard drives in our data centers. But, what does that even mean? How do you define a failure, and how does that definition define or obscure the realities of drive performance?
Traditional software-based security cannot prevent data destruction or unauthorized cloning when authorized credentials are compromised. This presentation introduces a hardware-level physical intervention approach with a zero-OS FPGA architecture to overcome vulnerabilities at the L3/L4 layer. This invisible Layer-2 transparent bridge, lacking IP or MAC addresses, performs sub-nanosecond-level gate-level Deep Packet Inspection (DPI) on the pipeline. When high-risk storage commands (deletion, unauthorized data transfer, etc.) are detected at the silicon level, the data flow is hardware-intercepted and held in a "Catch-and-Release" buffer until a physical human confirmation is received.
This presentation demonstrates how archive poisoning attacks compromise Retrieval-Augmented Generation (RAG) systems by embedding malicious instructions inside documents that influence model behavior during retrieval. Using a customized agentic testing workflow, attendees will see how a single poisoned document can manipulate outputs, exfiltrate sensitive information, bias threat intelligence results, and abuse trust relationships across a RAG-powered environment without modifying model weights.
Traditional software-based security cannot prevent data destruction or unauthorized cloning when authorized credentials are compromised. This presentation introduces a hardware-level physical intervention approach with a zero-OS FPGA architecture to overcome vulnerabilities at the L3/L4 layer. This invisible Layer-2 transparent bridge, lacking IP or MAC addresses, performs sub-nanosecond-level gate-level Deep Packet Inspection (DPI) on the pipeline. When high-risk storage commands (deletion, unauthorized data transfer, etc.) are detected at the silicon level, the data flow is hardware-intercepted and held in a "Catch-and-Release" buffer until a physical human confirmation is received.
For over 14 years, Backblaze has been collecting the failure rates of hard drives in our data centers. But, what does that even mean? How do you define a failure, and how does that definition define or obscure the realities of drive performance?