Glitching Edge Ai Fault Injection On Npus, Automated By LLM Agents
summary
Voltage glitching on a commercial Rockchip NPU corrupts object detection — hallucinations, phantom boxes and manipulated confidence scores occur without changing model or firmware. A lab rig with standard voltage fault-injection gear, EM probe and automated controller called 5Pi runs the campaigns; manual exploration yields about 4,700 experiments to map effects. An LLM-driven agent autonomously analyzes outputs, adjusts glitch delay/width/voltage and typically LLM agent finds red in ~1,400 attempts, cutting time and scaling the attack.
AI-written summary · about
chapters (8)
- 0:00Introduction to AI security
- 1:25Fault injection setup
- 5:27CNN fault analysis
- 10:29Developing an LLM agent
- 15:35Proof of concept demo
- 22:14Automating NPU testing
- 26:09Lessons and findings
- 27:21Conclusion and future work
related talks
21mGlitchGlück: Enabling Software Vulnerabilities through Guided Hardware Fault Injection
46mChip.Fail - Glitching the Silicon of the Connected World
43mImplementing Practical Electrical Glitching Attacks
43mRaiden Glitching Framework
31mA Dummy's Guide to Agentic Exploit Generation
01h 49mAI engineering with open access LLMs that lie, curse, and steal
39mGlitching AP4: A Technical Deep Dive Into Tesla’s Autopilot Computer
27mPMFault: Voltage Fault Injection on Server Platforms Through the PMBus