Hardware Hacking Evals

The goal is to quantify how well models can 'hack' hardware. This isn't as easy as testing them on pure software tasks - RE work often involves circuit modification and has a high risk of bricking the device under test. The plan would be

• Assemble a collection of peripherals and embedded devices, targeting popular + cheap ones that many consumers use (multiples of each device so that models start from ~stock)
• Build some basic physical 'test harness' that gives the models under test tools to, e.g., power cycle the device, (dis)connect it from USB, read the voltage level from a few test points... The goal is to test what is possible without extensive hardware modifications, but e.g. reading the status LEDs programmatically will replicate what could be done with a human in the loop
• Develop an eval based on compromising the hardware: modifying the firmware in some verifiable way and/or leveraging that to execute some significant action. For example, for a USB keyboard, models may score points by 1) modifying the device identifier (proof of minimal firmware mod) then 2) triggering a malicious script when the keyboard is plugged in

Because this is an under-publicized area, I think even a qualitative first pass with a few devices would have a significant impact re: public perceptions of capabilities, although a more thorough and quantitative eval is always nice to track progress with more granularity.