This article documents a novel, AI-driven hardware fault injection attack that successfully bypassed Secure Boot on an Espressif ESP32 SoC using voltage crowbar glitches. The attack was orchestrated entirely by the Claude Code AI agent, which controlled the ChipWhisperer Husky and lab supply hardware and wrote all necessary tooling. The vulnerability is specific to ESP32 Secure Boot V1, which has been mitigated in ESP32 V3 via a ROM code update, though the platform remains vulnerable to other fault injection methods.
Sunday, May 10, 2026 TL;DR:We usedClaude Codeto reproduce a Fault Injection attack whereSecure Bootis bypassed on an EspressifESP32SoC. We gaveClaudefull control to all the hardware tooling. All software tooling was written byClaudeusing third-party libraries. None of the code was written by humans. All this was created organically while glitches were being injected in the background (i.e., no downtime). As far as we know, this is the first publicly-documented AI-driven Fault Injection attack to date. This type of agentic workflow for finding/exploiting hardware vulnerabilities is likely here to stay, as it will be for software vulnerabilities. Day 1: Day 2: We reproduced a realFault Injection attack, whereESP32’sSecure Boot V1is bypassed using Voltage crowbar glitches, usingClaudedoing all the engineering. The AI did it all. From configuring theChipWhisperer Huskyand theRiden RK6006lab supply, to writing the attack script, debugging hardware quirks, reverse-engineering the bootROMwith a sub-agent, building a live monitoring dashboard while the campaign was running, and tuning parameters across thousands of glitch attempts. We supervised, asked questions, and looked at the dashboards it created for us. To our knowledge, this is the first publicly-documentedAI-driven Fault Injection attackat this depth. We’re sharing this not because theSecure Bootbypass itself is novel, that ground has been already covered in our previous blog posts, as well as in research conducted by others. What’s new here is the workflow, which is likely here to stay. Disclaimer:The attack described in this blog post is only applicable toESP32 V1and has since been mitigated byESP32 V3(see Espressif’sImpact Analysisfor more details). Note, the attack is mitigated by modifying the ROM code, theESP32 V3is still vulnerable to Fault Injection attacks (see our otherresearch). We simply kicked-off this adventure by tellingClaudewhat we were going to do. Me:“We are going to bypass ESP32’s Secure Boot using ChipWhisperer Husky.” Moreover, we told it to store everything it learns in a wiki format we like, which allows us andClaudeto remember what we are working on. Me:“Store and keep track of everything you learn using Hugo and the relearn theme.” Then we connected all the required tooling via USB to a freshly installed laptop with Ubuntu 26.04 LTS. Moreover, we giveClaudefull power using--dangerously-skip-permissions, as the laptop is not used for any other purpose. The hardware was wired up by hand. It’s the same setup we use for ourTAoFItraining.Claudewrote all tooling from scratch, relying on third-party libraries where possible. For example, we usedchipwhispererfor the Husky,picosdkfor the PicoScope,pyserialfor UART, an open-sourceRK6006-Python-Modulefor the lab supply. Here is a picture of the actual setup: The actual setup used for this AI adventure. Early on we askedClaudeto keep track of what it was building: Me:“While you learn about all the tooling; please keep updating a diagram of all the connections.” This resulted eventually in the following diagram: