The Tech Blog

Today let’s revisit an older piece of hardware security tech. I just wanted to illustrate that this tech is out there for informative purposes. A naive looking USB drive can fry your personal device.

The USB Killer appeared publicly around 2015. Early versions looked like ordinary USB Type-A drives, but instead of exchanging data, they used power from the port to create destructive electrical pulses. The idea later became a commercial stress-testing product, and the current generation is USBKill V4.

Fundamentally, the hardware has four blocks. It takes a low-voltage energy source, feeds a voltage-conversion stage, stores energy in capacitors, then switches that stored energy back into USB signal circuitry. Early designs charged the capacitors to around -110V, while later versions reached roughly -220V.

The interesting part is why normal USB current limiting on your host is not necessarily saving you. The host may limit how much current a device draws from 5V, but the device can accumulate that energy slowly and release it very quickly. Repeated pulses can over-stress ESD protection, damage the USB block in your PC, and potentially reach a hub, controller or SoC behind it.

USB-C made the original approach harder because a compliant Type-C source waits for a valid connection before enabling VBUS. Modern USBKill V4 versions can carry an internal battery, so they no longer need the target to supply that energy. Adapters also extend them to USB-C and other interfaces. Survival still depends on the host’s protection design.

Why does such hardware exist beyond malicious use as of today? Its role is destructive stress testing, helping hardware teams test whether ports survive abnormal electrical events and whether protection fails safely. It’s a tool that can be misused.

Your takeaway today should be to treat an unknown USB device as unknown hardware, not just unknown storage. A random device can severely damage a port or motherboard before software security gets any chance to react. You need to know that this exists.

GPT-6(Astra) launched this week. It’s an insane step-up model. But one small part of the launch caught my attention more than the benchmarks. OpenAI showed Astra operating KiCad, taking a schematic, placing components and routing the PCB as well as a great intern would. We are on the cusp of massive takeoff.

I wrote about AI in electrical engineering almost two years ago, and my view has only become stronger. Over the next couple of years, I think conventional embedded product design will become heavily AI-assisted and to a great extent automated.

Give the system a product specification and it should be able to work backwards from it. What processor fits? Which sensors, regulators and interfaces make sense? What is in stock? Which alternatives reduce cost or supply-chain risk? It can read datasheets, compare trade-offs, build the architecture, create the schematic and eventually move into placement, routing and verification. Today’s models already help with several of these steps. I personally use them and it’s not too shabby.

After more than a decade working on embedded products, I know how much time goes into repetitive work: comparing parts, reading datasheets, creating symbols and footprints, checking reference circuits, updating BOMs and layouts. I think much of that will move to agents. Your value as an engineer shifts towards defining requirements, understanding trade-offs, spotting failure modes, validating the design and deciding what should actually be built.

I find that exciting. Most work will get displaced. The useful response is to learn these tools, change how you work and keep moving up the abstraction stack. Continuously upskill, or else you will be left behind.

My guess is that within 2-3 years, giving an AI a complete enough specification and getting most of a conventional embedded board back for engineering review will feel normal.

What do you folks think? Is 2–3 years too aggressive? Personally, I don’t think so. If anything, remember that the KiCad demo we are seeing today is likely the least capable version this technology will ever be.