I find it charming that to distinguish Fil-C from the K&R language they use the term 'Yolo-C'. I have never used inline asm before, I actually didn't realise how much behaviour it's specifying! (When I've needed asm it was on non-gcc compilers)
I wonder if an adversarial user could bypass the checks and achieve memory corruption / code execution. Maybe not a practical attack in most situations but a fun exercise.
> This includes things like asm volatile("" : : : "memory"), which is an old-school way of saying atomic_signal_fence(memory_order_seq_cst).
Not quite. AIUI, the first is just a barrier for the compiler, while the second is also a CPU memory barrier. Godbolt seems to confirm that.
What is more frightening about this than safe C assembly is that this level of implementation is achievable not with a SOTA model, but with a cost effective model like KIMI. There was human judgment involved in the middle, but reading the article, My reading of the process is as follows:
1.A developer identified the necessity of inline assembly.
2.Defined the safety boundaries for 'memory-safe' inline assembly.
3.Established strict policies for memory access.
4.Curated an allowlist of permissible instructions.
5.Set rigorous test criteria and 'done' conditions.
In short, with the overall guardrails in place, a sub agent loop was run, and this level of code was produced. This raises a number of interesting points about how we should use AI. I haven't looked at all the code, but the idea of passing assembly through safe zones without memory access, and using that as a foundation to achieve this level of implementation through AI, is quite impressive
I find it charming that to distinguish Fil-C from the K&R language they use the term 'Yolo-C'. I have never used inline asm before, I actually didn't realise how much behaviour it's specifying! (When I've needed asm it was on non-gcc compilers)
I wonder if an adversarial user could bypass the checks and achieve memory corruption / code execution. Maybe not a practical attack in most situations but a fun exercise.
> This includes things like asm volatile("" : : : "memory"), which is an old-school way of saying atomic_signal_fence(memory_order_seq_cst).
Not quite. AIUI, the first is just a barrier for the compiler, while the second is also a CPU memory barrier. Godbolt seems to confirm that.
https://godbolt.org/z/a844zKej8
Your godbolt code used atomic_thread_fence
The quote uses atomic_signal_fence.
If you find a way to bypass my checks, file a bug. I tried very hard to break it. My agent loops tried even harder
Oops, you're right. I was thinking of those as nearly interchangeable but they're actually pretty different.
I might give it a try when I have a chance, I'll let you know if anything comes of it.
What is more frightening about this than safe C assembly is that this level of implementation is achievable not with a SOTA model, but with a cost effective model like KIMI. There was human judgment involved in the middle, but reading the article, My reading of the process is as follows:
1.A developer identified the necessity of inline assembly.
2.Defined the safety boundaries for 'memory-safe' inline assembly.
3.Established strict policies for memory access.
4.Curated an allowlist of permissible instructions.
5.Set rigorous test criteria and 'done' conditions.
In short, with the overall guardrails in place, a sub agent loop was run, and this level of code was produced. This raises a number of interesting points about how we should use AI. I haven't looked at all the code, but the idea of passing assembly through safe zones without memory access, and using that as a foundation to achieve this level of implementation through AI, is quite impressive