AMD Safe RET cracked by interrupt timing attack

---

AMD Safe RET mitigation bypassed via interrupt timing attack on Zen 1-4 processors. A MIT CSAIL researcher demonstrated that Linux's Safe RET implementation can be weakened through precise interrupt injection. The vulnerability affects AMD EPYC 7001 through 9004 series servers spanning 2017-2024 deployments. Mitigation patches merged to mainline kernel on August 7.

This is the latest reminder that speculative execution mitigations themselves become attack surfaces. Six years past Spectre, the threat model keeps evolving.

In this issue:

• AMD Safe RET bypass: interrupt timing attack, kernel patches merged
• OVHcloud Exten NVMe storage: Reed-Solomon 4+2 encoding, 1.5x overhead
• dbt Labs cuts AI token costs 98% with warehouse-based context preprocessing
• Asahi Linux Apple NVMe driver: M4 firmware compatibility patches

---

AMD Safe RET Mitigation Weakened by Interrupt Timing Attack

The Signal

Linux's Safe RET mitigation for AMD Zen 1-4 SRSO vulnerability bypassed via precise interrupt injection.

What Changed

• MIT CSAIL researcher demonstrated interrupt-based attack weakening SRSO mitigation on Zen 1-2 processors.
• Attack potentially affects Zen 3-4 processors, though not yet demonstrated in practice.
• Vulnerability stems from Linux kernel's implementation of Safe RET, not AMD silicon design.
• Mitigation patches merged to mainline kernel today, per AMD Security Bulletin SB-7011.
• Affected processors span 2017-2024 AMD server deployments (EPYC 7001 through 9004 series).

Operational Impact

Teams running AMD Zen-based servers need kernel updates to close this mitigation gap. The attack requires local code execution to inject precisely-timed interrupts, making it primarily relevant for multi-tenant environments. This demonstrates that speculative execution mitigations themselves become attack surfaces—six years past Spectre, the threat model continues evolving.

Watch For

Similar timing-based attacks may target other speculative execution mitigations across architectures. Monitor whether performance overhead increases with the hardened Safe RET implementation.

---

OVHcloud's NVMe Block Storage Engine

The Signal

OVHcloud detailed Exten, their production NVMe block storage engine running 1+ years, built with Reed-Solomon 4+2 encoding.

What Changed

• Six-server clusters use NVMe-oF protocol for both volume exposure and inter-host communication.
• Data plane in C++, control plane in Go, communicating via gRPC.
• Reed-Solomon 4+2 encoding achieves 1.5x storage ratio versus 3x for triple mirroring.
• Raft consensus handles control plane state; user IO bypasses Raft entirely.
• Data-centric processing uses pointer manipulation to avoid buffer copies from NIC to drive.

Operational Impact

This replaces OVHcloud's Ceph deployments with a purpose-built stack optimized for NVMe performance. Reed-Solomon encoding cuts storage overhead by 50% compared to mirroring while tolerating two drive failures. The control plane can survive two host failures in a six-node cluster. Teams evaluating distributed block storage should note the architectural split: synchronous Raft for metadata, asynchronous zero-copy IO for data.

Watch For

OVHcloud is exploring RDMA transport to replace TCP for lower latency. They're soliciting feedback on features, so expect the architecture to evolve beyond Ceph feature parity.

---

dbt Labs Cuts AI Token Costs 98% With Warehouse-Based Context Layer

The Signal

dbt Labs reduced AI token costs 98% by pre-processing Gong transcripts in their warehouse instead of direct MCP connections.

What Changed

• Direct Claude-to-Gong MCP consumed 50,000 tokens ($0.25) per transcript analysis with hundreds of daily users
• Warehouse-based summarization using native AI functions reduced volume 20x, tokens to hundreds per call
• Pattern applies to Salesforce emails, Zendesk tickets, Jira descriptions already replicated as TEXT columns
• Fivetran now replicates unstructured files (PDFs, audio) and SaaS attachments to Snowflake/BigQuery/Databricks file storage
• dbt orchestrates Snowflake Cortex AI_COMPLETE, AI_EMBED functions as SQL-based transform models

Operational Impact

Direct MCP connections create compounding costs: vendor API charges plus token fees on every query. With 500M+ tokens of source data and repeated user queries, costs scale linearly. Warehouse pre-processing inverts this: read raw data once in batch, summarize with native AI functions, serve summaries to all agents. Teams running Claude against CRM or support data should evaluate whether existing warehouse text fields can replace real-time MCP calls. The cost model shifts from per-query to per-ingestion.

Watch For

This positions data teams as AI context owners, not just BI pipeline builders. The pattern requires warehouse platforms with built-in AI functions—Snowflake Cortex, BigQuery AI, Databricks ML—and incremental models to keep context current as source data changes.

---

Apple NVMe Linux Driver Patched for M4 Firmware Compatibility Issues

The Signal Asahi Linux patched its Apple NVMe driver after macOS 15+ firmware caused kernel panics and init failures.

What Changed

• New NVMe firmware shipped with macOS 15 is stricter than previous versions.
• M4 and later controllers require page-aligned admin queue buffers.
• Patches fix kernel panic on controller initialization failure.
• Driver now works around missing chicken bit in newer firmware.
• Firmware update is prerequisite for M4 Linux support on Apple Silicon.

Operational Impact Asahi Linux users who upgraded to macOS 15 or later hit controller init failures until these patches land. The M4 buffer alignment requirement shows Apple's firmware continues to impose undocumented constraints. Teams reverse-engineering Apple hardware face ongoing maintenance as each macOS release tightens firmware behavior without public documentation.

Watch For Expect similar breakage patterns as M4 Pro, Max, and Ultra support progresses. Apple firmware updates remain a moving target for non-macOS kernel developers.

---

Quick Reads

• AMD Safe RET — Linux mitigation for Zen SRSO vulnerability bypassed via precise interrupt injection.
• OVHcloud Exten — Production NVMe storage engine uses Reed-Solomon 4+2 encoding, cuts overhead 50%.
• dbt Labs — Warehouse-based AI context pre-processing slashed token costs 98% versus direct MCP.
• Apple NVMe Driver — Asahi Linux patched for macOS 15 firmware compatibility on M4 controllers.