Study shows AI code review cannot detect GNU strip-based supply chain backdoors

A new research paper (arXiv 2607.24888) demonstrates that Ken Thompson's classic 'trusting-trust' compiler attack can be replicated using GNU strip, a build utility that processes only binary ELF files rather than source code. In the NixOS bootstrap environment, a single malicious strip binary injected into the seed can propagate a hidden payload across build generations, persisting even after the original infected binary is removed. Because AI code review tools analyze source diffs and patches, they have no visibility into the binary dependency chain where the malicious code actually resides. The authors argue this is a fundamental limitation, not one that more powerful AI models can overcome, since the threat exists entirely outside the source layer. Teams using AI review are advised to apply it strictly to source-level analysis and rely on separate artifact provenance and build reproducibility measures to verify what actually runs in production.