Comment by vlovich123
3 days ago
Reflect in what way? The primary focus of that talk is that it’s possible to infect the binary of a compiler in a way that source analysis won’t reveal and the binary self replicates the vulnerability into other binaries it generates. Thankfully that particular problem was “solved” a while back [1] even if not yet implemented widely.
However, the broader idea of supply chain attacks remains challenging and AI doesn’t really matter in terms of how you should treat it. For example, the xz-utils back door in the build system to attack OpenSSH on many popular distros that patched it to depend on systemd predates AI and that’s just the attack we know about because it was caught. Maybe AI helps with scale of such attacks but I haven’t heard anyone propose any kind of solution that would actually improve reliability and robustness of everything.
[1] Fully Countering Trusting Trust through Diverse Double-Compiling https://arxiv.org/abs/1004.5534
The proposed solution seems to rely on a trusted compiler that generates the exact same output, bit-for-bit, as the compiler-under-test would generate if it was not compromised. That seems useful only in very narrow cases.
You have a trusted compiler you write in assembly or even machine code. You then compile a source code you trust using that compiler. That is then used for the bit for bit analysis against a different binary of the compiler you produced to catch the hidden vulnerability.
It's assumed that in this scenario you don't have access to a trusted compiler; if you do, then there's no problem.
And the thesis linked above seems to go beyond simply "use a trusted compiler to compile the next compiler". It involves deterministic compilation and comparing outputs, for example.
1 reply →
I believe the issue is if an exploit is somehow injected into AI training data such that the AI unwittingly produces it and the human who requested the code doesn't even know.
That’s a separate issue and specifically not what OP was describing. Also highly unlikely in practice unless you use a random LLM - the major LLM providers already have to deal with such things and they have decent techniques to deal with this problem afaik.
If you think the the major LLM providers have a way to filter malicious (or even bad or wrong) code from training data, I have a bridge to sell you.
1 reply →