Cedric Clyburn explains on IBM Technology how AI-written code expands an application’s software supply chain and increases pressure on review.

Treat the agent as another supplier

  • Production depends on packages, transitive dependencies, container images, build pipelines, secrets and generated code.
  • Dependency bugs, poisoned package registries and compromised builds create different entry points.
  • AI increases code volume and can accelerate vulnerability discovery, making verification capacity more important.

Four controls to put near the work

  1. Generate an SBOM, a software bill of materials, so teams can locate affected packages and versions quickly.
  2. Automate dependency, secret and authenticity checks; keep changes small enough to review.
  3. Review generated code independently, using a separate session or model alongside tests and checks.
  4. Sign artifacts and record build provenance, using SLSA, a framework for software supply-chain integrity, as a guide.

Plan for fixes that cannot wait for a major upgrade

  • Clyburn names Akrites and Lightwell as efforts addressing open-source vulnerability response.
  • Their first-party descriptions distinguish Akrites’ coordinated disclosure and remediation from Lightwell’s patched dependencies and backports.
  • Caveat: separating the reviewer can reduce shared context, but the talk provides no measured assurance that a second AI reviewer catches every defect.

“If something were wrong, how would we know?”