Black Hat · November 19, 2026

GPUBreach: include GPU memory and driver trust in isolation reviews

GPUBreach: include GPU memory and driver trust in isolation reviews video thumbnail
Why it matters

The research behind this Black Hat talk shows how GPU Rowhammer can extend beyond model-data corruption. On an RTX A6000 with GDDR6 memory, researchers corrupt GPU page-table entries to obtain cross-process memory access, demonstrate key and model-data leakage, and alter GPU code. They then use GPU-accessible driver buffers to trigger host-driver memory-safety flaws and obtain root privileges while the IOMMU remains enabled. The IOMMU still restricts DMA addresses; the escalation abuses trusted state inside permitted buffers. The demonstrated Rowhammer chain assumes ECC is disabled and requires Unified Virtual Memory. Tests on additional GPUs used simulated bit flips for the driver stage, so they do not establish physical Rowhammer susceptibility on those devices. NVIDIA recommends enabling system-level ECC where supported, while the researchers caution against treating ECC as a universal guarantee. This is controlled research with a public artifact, not evidence that all GPUs or cloud isolation modes are exploitable.

My takeaway: Record GPU model, memory type, driver version, ECC state and UVM availability in your infrastructure inventory. Restrict untrusted CUDA execution and review tenancy boundaries. Keep IOMMU enabled, verify supported ECC settings, and evaluate driver hardening separately from memory-error mitigations.
Keep exploring

More curated notes connected through Adversarial ML and AI Red Teaming.

OpenAI News · framework

Pacing model development in an era of cyber-critical capabilities

OpenAI says preliminary evidence that Astra may meet its Critical cybersecurity threshold led it to pause frontier reinforcement-learning work for two weeks and keep its largest planned run on hold. New safeguards include stronger workload and network isolation, continuous boundary testing, token-level monitoring that escalates suspicious tool activity, and broader alignment checks for deception, reward hacking, and unauthorized access.

OWASP GenAI Security Project · guide

OWASP Top 10 for Agentic Applications for 2026

OWASP's community guide organizes agentic-system risk into ten categories, including goal hijacking, tool misuse, identity and privilege abuse, memory poisoning, insecure inter-agent communication, cascading failures, and rogue-agent behavior. It provides a shared taxonomy and mitigation starting point rather than a certification checklist or evidence that a deployed system is secure.

OpenAI News · framework

A blueprint for democratic governance of frontier AI

OpenAI proposes a three-part U.S. frontier-AI governance model: harmonize emerging state safety laws into a federal baseline, strengthen CAISI as an evaluation and standards institution, and coordinate a broader resilience program. Proposed controls include severe-risk evaluations, transparency reports, independent audits, safety-incident reporting, model-weight security, whistleblower protection, and periodic technical assessments.