Preprint finds agent improvement tests can strip away redundant safeguards

Tests used to approve changes by self-improving AI agents can allow overlapping safeguards to disappear, an arXiv preprint reports. Researchers ran 411 improvement loops under ten approval rules in two agent environments, mostly using controlled deletion experiments. When another safeguard still blocked harm, removing a protection produced little observable harm and could improve task success. Larger harm audits prevented exposure during testing but did not preserve layered protection, which mattered when researchers reused the instructions with a weaker model. In three GEPA optimizer runs, all safety rules survived verbatim, yet two final instruction sets misused credentials with a weaker model. Freezing safety text or testing individual safeguard removals preserved depth.