1. Failure to differentiate data from instructions
2. Business logic flaws
3. Insecure operating environment
4. Incorrect use of cryptography
I mean, I'm kinda kidding, but #1 is #1.
Execute without read is not supported so much though.
Aside from language level enforcement (Java/Rust/...)
It's what still tripping half the people working on security when it comes to LLM.
1. Error that causes security problems
What is being secured (as in, whose interests), and from who?
Security is all nice until one day you find yourself being seen as threat actor. See e.g. all the modern computing devices you "own".
OWASP already covers this
It's almost always memory safety violations. The rest are basically application-level bugs, like not properly checking inputs.
CVEs are a big deal, but humans are the real vulnerability
I'd phrase that as "know your data access ordering," and include races.
"escape your outputs" works for all data and is actually achievable. It means even escaping data that was never an "input" from the outside the system, even developer-created data, which absolutely can (and does) cause problems to ignore.
"Can't do math" (off-by-1 errors etc)
"Sanitize your inputs" (escape sequences, parsing etc)
"What are you referencing?" (null pointers, use-after-free etc)
and error handling. Plain error codes bubbling up a call stack, or bypassing that somehow which may violate program structure.
How many of these fundamental issues are there?