How We Taught ChatGPT-4 to Break mbedTLS AES With Side-Channel Attacks
summary
Shows GPT‑4 generating device‑under‑test and driver C code, building a leakage model and orchestrating USB passthrough to cloud oscilloscopes (ChipWhisperer / picoscope) over a KVM REST API, then running Correlation Power Attacks against mbedTLS AES on STM32F3. The system recovers a key byte in seconds (demo ~12s), reuses the exact same code across six STM32 variants, and auto‑fixes build errors, while facing practical limits: hallucinations, 4k token context size and large binary handling that require token optimization and asynchronous APIs.
AI-written summary · about
chapters (18)
- 0:00Introduction and agenda
- 1:17GPT for security
- 3:58Side-channel attack basics
- 7:44Live demo: simple attack
- 8:39IoT security challenges
- 10:24Integrating GPT with hardware
- 12:37Technical setup for demo
- 14:49Prompt engineering examples
- 17:28GPT capabilities in security
- 19:42Context and hallucination
- 22:53Token optimization techniques
- 23:52Managing LLM limitations
- 26:58Handling request timeouts
- 28:27Live demo: advanced analysis
- 30:17Live demo: WolfSSL attack
- 32:33Live demo: error correction
- 33:33Conclusions and Q&A
- 35:59Audience questions
related talks
49mChatGPT: A New Generation of Dynamic Machine Based Attacks?
19mReality Check on Side-Channels: Lessons learnt from breaking AES on ARM Cortex-A72...
51mExploiting Hardware Glitches And Side Channels In Perfect Software
01h 07mPEW PEW PEW: Designing Secure Boot Securely
59mBreaking AES with side channel analysis
44mSecuring LLM-Powered Apps: Overcoming Security and Privacy Challenges
01h 02mCryptographic Side-Channels on Embedded Devices
24mSecurity through Transparency: Tales from the RP2350 Hacking Challenge