What's this about?
Two papers, opposite ends of the same pipeline. CamoDocs poisons retrieval-augmented generation with camouflaged documents: RAG pulls from public or user-editable sources, so an attacker can inject documents crafted to steer outputs while reading as ordinary content. GraftyVul synthesises insecure programs by grafting real-world vulnerabilities into clean code, because existing vulnerability datasets always sacrifice at least one of realism, scale or label quality.
Why is building a vulnerability dataset hard?
You want three things and can normally have two. Real bugs from real projects are realistic but few, and their labels are argued about. Synthetic bugs scale infinitely and have perfect labels but look nothing like production code, so detectors trained on them fail in the wild. Grafting is the attempt at all three: take a real vulnerability with a known mechanism, transplant it into a different real codebase, and you inherit realism from both while keeping the label exact.
Why does this matter to you?
- GraftyVul is directly usable if you ever need to test whether a scanner detects what it claims to.
- CamoDocs is the poisoning counterpart to ContextLeak: one gets bad data in, the other gets good data out.
- Both underline that the interesting attacks now target the data layer, not the model.