AI Coding Assistants Rarely Verify Software Supply-Chain Signals Before Installing

Summary

A preregistered study audits whether AI coding assistants check software supply-chain trust signals before installing research software. The researchers used six open-source projects, three from high-performance computing and three from quantum computing, selected from an 87-project corpus. For each project, they created modified copies representing the absence of a signal, four individual signal classes, two signals from the wrong issuer, all four signals together, and documented conflicts in the project’s metadata. The experiment produced 1,920 registered trials involving three models and two assistant operating setups, with conditions both including and excluding an approval step; a supplement also covered three frontier models. The researchers scored behavior from container logs and recorded the cost of every trial rather than relying on assistant statements. Verification occurred in only 9 of 1,920 registered trials, or 0.5%; no control trial involved opening a provenance signal, and no trial ran a verification command. The presence of a trust signal therefore produced no measurable change in observed behavior. The paper concludes that publishing machine-checkable signals is necessary but insufficient, that higher price did not ensure verification, and that verification should be built into the program running the assistant. The authors release the trial-level cost ledger, protocol, and logs.