Skills
LLM-generated services pass their tests and then leak memory for 48 hours straight
arXiv 2608.26391 (2026-08-26) ran generated JavaScript, Python and Rust applications for 48 hours while tracking memory, response time and throughput, and found memory growth with statistically significant upward trends in most systems. The authors' point is narrow and useful: functional correctness says nothing about operational reliability for anything that runs continuously, so a long-running soak belongs in the acceptance check for generated services. They also saw the same aging in human-written implementations, so this is a testing gap rather than an indictment of generated code.
↳ Follow the thread