PROJECT SHADOWCORE — GPU Firmware Security Research

Overview

Security research project investigating whether code can persist within GPU firmware (specifically NVIDIA's GSP — GPU System Processor), surviving driver reloads and potentially reboots, in a location invisible to host-side EDR, antivirus, and forensic tooling.

Goal: Demonstrate the attack surface exists, then build the detection and defense tooling that doesn't exist yet. All findings will be responsibly disclosed to NVIDIA PSIRT and MITRE.

Responsible Disclosure

This is authorized security research conducted for the purpose of:

1. Proving whether GPU firmware persistence is feasible (even a "hello world")
2. Building detection tooling for a currently unmonitored attack surface
3. Responsible disclosure to GPU vendors (NVIDIA, AMD, Intel)
4. Academic publication after vendor coordination

No malicious use. The PoC payload is a proof of execution (register write, hello world), not a weaponized implant.

Project Structure

PROJECT-SHADOWCORE/
├── docs/                          # Research documentation
│   └── PHASE0-RESEARCH-PLAN.md    # Detailed Phase 0 plan & requirements
├── tools/                         # Research tooling
│   ├── firmware-extract/          # GSP firmware blob extraction & triage
│   ├── bar-mapper/                # PCI BAR region mapping
│   ├── vram-persistence/          # VRAM persistence testing suite
│   └── gsp-rpc-monitor/          # GSP RPC interface cataloger
├── analysis/                      # Analysis outputs
│   ├── firmware-blobs/            # Extracted firmware binaries
│   ├── ghidra-projects/           # RE project files
│   └── notes/                     # Research notes
├── detection/                     # Defensive tooling (Phase 3)
│   ├── yara-rules/                # Detection signatures
│   ├── integrity-checker/         # Firmware integrity verification
│   └── forensics/                 # GPU memory forensics
├── scripts/                       # Utility scripts
├── tests/                         # Test suites
└── requirements.txt               # Python dependencies

Phase 0 Tools

Firmware Extraction & Triage

# Scan system for installed NVIDIA firmware blobs
python tools/firmware-extract/extract_gsp_firmware.py --scan-system

# Analyze a specific firmware blob
python tools/firmware-extract/extract_gsp_firmware.py /path/to/gsp_ad10x.bin -o analysis/firmware-blobs/

PCI BAR Mapping

# Auto-detect NVIDIA GPUs and list BAR regions
sudo python tools/bar-mapper/map_gpu_bars.py

# Probe register ranges (requires root)
sudo python tools/bar-mapper/map_gpu_bars.py --probe

# Target specific GPU
sudo python tools/bar-mapper/map_gpu_bars.py --bdf 01:00.0 --probe

VRAM Persistence Testing

# Write test patterns to VRAM
sudo python tools/vram-persistence/test_vram_persistence.py write --bdf 01:00.0

# After triggering a state transition (driver reload, suspend, etc.):
sudo python tools/vram-persistence/test_vram_persistence.py verify --bdf 01:00.0

# Automated driver-reload test
sudo python tools/vram-persistence/test_vram_persistence.py test-driver-reload --bdf 01:00.0

GSP RPC Interface Cataloger

# Clone NVIDIA source first
git clone https://github.com/NVIDIA/open-gpu-kernel-modules /opt/nvidia-open

# Generate RPC catalog
python tools/gsp-rpc-monitor/gsp_rpc_catalog.py /opt/nvidia-open
python tools/gsp-rpc-monitor/gsp_rpc_catalog.py /opt/nvidia-open --format json -o analysis/notes/rpc_catalog.json

Hardware Requirements

• Test GPU: NVIDIA RTX 3050+ (Ada Lovelace preferred for RISC-V GSP)
• Display GPU: Separate GPU or integrated graphics for display output
• OS: Linux (Ubuntu 22.04+ or Fedora 38+), kernel 6.1+
• Access: Root privileges for MMIO/BAR access and driver management

Software Dependencies

# System packages
sudo apt install pciutils binwalk ghidra linux-headers-$(uname -r)

# Python environment
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

# NVIDIA open-source kernel modules (for reference, not installation)
git clone https://github.com/NVIDIA/open-gpu-kernel-modules /opt/nvidia-open

# envytools (NVIDIA RE toolkit)
git clone https://github.com/envytools/envytools /opt/envytools
cd /opt/envytools && cmake . && make

Research Phases

Phase	Focus	Status
Phase 0	Reconnaissance — map attack surface	Active
Phase 1	Firmware RE — understand GSP internals	Planned
Phase 2	PoC — prove code execution (hello world)	Planned
Phase 3	Detection — build defensive tooling	Planned

Key References

• NVIDIA Open-Source Kernel Modules
• envytools — GPU RE toolkit
• Nouveau Driver — Open-source NVIDIA driver
• Falcon ISA Docs — Microcontroller ISA
• Vasiliadis et al. (2010) — "GPU-Assisted Malware"
• "Jellyfish" (2015) — VRAM-resident rootkit PoC

License

This research is conducted under responsible disclosure principles. Tools in this repository are for authorized security research only.