Researchers have audited the behavior of three AI systems designed to evolve and improve their performance over time. The study found that while these systems can become more capable in certain tasks, they also increase their exposure to potential attacks and unauthorized financial state changes. One system, ReasoningBank, was able to raise its utility without increasing its attack success rate. However, the other two systems showed a trade-off between capability gains and security drift. The study highlights the need for auditing self-evolving AI systems beyond just measuring their accuracy. --- Why it matters: This research is important because it shows that simply measuring the performance of self-evolving AI systems may not be enough to ensure their security and reliability. Engineers working on these systems need to consider how they can prevent regressions, minimize attack surfaces, and maintain compatibility with existing tools and execution environments. Source: https://arxiv.org/abs/2608.17684