1. Look in the manual for statements of the form "Don't do X"
2. Do X
The idea is: stipulate that no attacker can ever inject Javascript into a browser. Assume we solve that problem completely. Now, how secure are DOM-based applications? Turns out: not that much more secure. Lots of very clever examples.
(Also, anything icamtuf touches is probably going to be good.)
Understanding the function, and boundaries, and interactions of a system, application, protocol, device et al and being able to identify who are threat actors, what the assets to protect are, and what the real threats against the entity are will point you at the more valuable things to start. Far, far too many people gloss over threat modeling because they don't equate it with technical work.
But on top of being "a great way to use all your skills up to this point" I think it can also be a strong motivation for picking up new ones.
Learning C super well is important, so is assembly, so is other topics but I think fundamentally, step one needs to be: "I have product X. What is it trusting? What is it trying to protect? How do I get around it?" Then, find similar vulnerabilities and continue on stumbling to the ultimate goal.