GPUThor breaks the error correction Nvidia recommended against Rowhammer
THE SO WHAT
GPUThor punching through ECC on Nvidia GPUs turns Rowhammer from a theoretical GPU risk into a live one—model hosts now have to assume memory corruption attacks can ride on their accelerator fleet. If you’re serving sensitive workloads on shared GPUs, treat isolation, firmware patching, and anomaly detection as urgent, not optional.
READ THE SOURCE
MORE FROM THE WIRE
Deep & Emerging TechGuest Post: Quantum Computing’s Next Phase May Be Driven by Compounding Breakthroughs
The quantum narrative is shifting from single headline breakthroughs to compounding, incremental gains — more like semiconductor roadmaps than moonshots. For operators, that means tracking vendor roadmaps and ecosystem standards now, even if production workloads are still several cycles out.
Deep & Emerging TechBlackstone Bet on Chips Drove a Debt Boom That’s Upending Credit
Private credit is being rewired around AI infrastructure — GPU-backed deals are now a mainstream asset class. If you’re building anything compute-intensive, your financing options just expanded, but so did scrutiny on utilization, power contracts, and resale value of your hardware.
Deep & Emerging TechAuthentication does not equal authorization: Are we asking passkeys to prove too much?
Passkeys solve “who are you,” not “what should you be allowed to do” — collapsing those layers is how over-permissioned accounts and lateral movement happen. CISOs should treat passkeys as one control in a broader authorization and policy engine, not as a reason to relax least-privilege design.
Deep & Emerging TechProtecting Trade Secrets In The Age Of AI
AI agents turn every employee device into a potential exfiltration channel for trade secrets — often via “helpful” assistants. If you haven’t updated confidentiality policies, DLP rules, and vendor contracts for AI-specific flows, assume your IP perimeter is already porous.