Hello guys. This is Lakin Mohapatra. Working as associate team lead with 7.5+ years of development exp. I am planning to write posts on the following topics :.- PHP.- Golang.- Rust.- Javascript.- Blockchain.- Cyber security. Follow me @lakincoder if I deserve your attention.

The more I watch South Indian films, the harder it gets to ignore the gaps in Odia filmmaking. No fresh direction. Weak screenplays. Few bold experiments. Odia cinema has talent — but where’s the vision?.We need better storytelling. We deserve better cinema. #OdiaCinema #Odia.

𝗙𝘂𝗻𝗱𝗮𝗺𝗲𝗻𝘁𝗮𝗹𝘀 𝗺𝗮𝘁𝘁𝗲𝗿. 𝗕𝘂𝘁 𝘀𝗼 𝗱𝗼𝗲𝘀 𝘆𝗼𝘂𝗿 𝗲𝗻𝘁𝗿𝘆 𝗽𝗼𝗶𝗻𝘁. 𝗦𝗼 𝗱𝗼𝗲𝘀 𝗺𝗮𝗿𝗸𝗲𝘁 𝗰𝘆𝗰𝗹𝗲. 𝗦𝗼 𝗱𝗼𝗲𝘀 𝘃𝗮𝗹𝘂𝗮𝘁𝗶𝗼𝗻.

Ever faced unexplained accuracy drops or broken models in prod?. 🔁 RT this thread to keep your team aware. Follow @lakincoder for more on AI safety, infra risks, and secure deployment in the LLM era.

RowHammer is back. But now it speaks CUDA. And it’s targeting AI — the very thing we’re trusting with critical decisions. Don’t let silent memory flips take your models down.

Want to go deeper?. Look up:. - GPUHammer (2025 paper). - Half-Double. - TRRespass. - DRAM Row Buffer Conflicts. - AI Model Robustness to Bit-Flips. - ECC limitations in consumer GPUs. You’ll realize: the future of AI needs hardware-aware security.

For devs, ML folks, and infra teams:. If you're running shared GPU workloads or deploying AI models into production…. - Treat model integrity like critical infrastructure. - Assume memory can lie. - Design systems with resilience to physical-level faults.

We’ve hardened code. We’ve sandboxed processes. We’ve patched vulnerabilities. But our AI models live in DRAM. And DRAM can be flipped — just by looking at it the wrong way. The enemy now? Physics. Think beyond software security.

TL;DR. - RowHammer flips bits. - GPUHammer does it on AI systems. - No logs. No alerts. Just broken models. - Accuracy silently drops. - And we can’t fix it with software alone. Time to take memory-level security seriously.

RowHammer isn’t over. It evolved. And in 2025, attackers have learned to weaponize GPU-based memory corruption. This hits:.Cloud infra. LLM-as-a-service providers. On-device models. Autonomous systems. AI-powered security tools.

In an AI-first world, memory corruption = AI corruption. Your model can hallucinate, misclassify, or fail not because of bad code — but due to bit flips induced by another workload. This is silent. And that's the danger.

🔁 Use redundancy + checksumming when critical inference is deployed.

What can you do as a developer, architect, or ML engineer?. Start here . ❌ Avoid running sensitive AI workloads on shared/public GPUs. ✅ Validate models post-training against known checkpoints. 🧠 Stay updated on GPU memory attacks like GPUHammer & Half-Double.

This isn’t a software bug. It’s a physics problem in how DRAM works. The electric charge used to access one row slightly bleeds into neighboring ones — and that’s enough to flip 0s to 1s. Modern DRAM is faster, but also more vulnerable.

Real-world GPUHammer attack flow:. You're running an AI job on a shared cloud GPU. Another user on the same card runs hammering code. Bit flips corrupt your model weights in DRAM. You save the model, deploy it. Now it's predicting garbage in prod — silently. Nightmare fuel.

What’s at risk?. - AI model weights during training. - In-memory authentication data. - Kernel or runtime configurations. - Control policies in edge devices. - Real-time vision models in self-driving or robotics. All of it can be flipped. No logs. No warning. Just chaos.

And here’s the scary part:. An attacker doesn’t need root. Just aggressive GPU code — like fast matrix multiplications or shader programs. From another tenant on a shared GPU. → Your AI workload is quietly corrupted. → They never even touch your app.

Imagine this:. You’re training a ResNet model. All’s going well. But a few bits get flipped in memory…. Accuracy drops from 76% to 12% — with no changes in code or logs. No errors. No crashes. Just broken models. That’s what GPUHammer can do.

Originally, RowHammer was a CPU-side attack. But now, in 2025, researchers have triggered the same behavior on NVIDIA GPUs — the same ones running:. - AI training. - LLM inference. - Cloud pipelines. - Video rendering. Meet: GPUHammer. This is next-level serious.

What is RowHammer?. Your system memory (DRAM) is structured like a giant spreadsheet — rows + columns. If one row gets “hammered” (accessed rapidly), nearby rows can experience electrical interference…. …which flips bits in other data. Yes, data you didn’t touch. 😳.

RowHammer is back — and this time, it’s targeting your GPUs and AI models. It’s no longer “just an old CPU bug from 2015.”. It’s a hardware-level attack that can flip your bits, poison your weights, and crash your systems — invisibly. Let’s break it down 👇.