Open-source AI agents that probe and attack websites by themselves have spread fast since one popular tool, ARTEX, was taken down and copied in many forks.

Cloudflare tracked more than 50 families of these tools across its network and set out six defences, including firewall rate limits, bot defence and identity-gated access to sensitive apps.

Developers wanted to stop surprise bills, and alerts alone could not stop runaway usage.

A beta feature for Workers and Durable Objects that pauses paid activity when a limit is hit. It is fed by a two-stage pipeline that adds up usage in one-minute and then five-minute windows.

Cuttlefish, Android's accurate virtual device, ran only on Linux, so Meta's Mac-based developers had to use a less accurate emulator alongside it.

Meta ported the QEMU path to macOS. It made vhost-user build outside Linux, raised memory-region slots in Apple's hypervisor backend, and added a native display and code signing.

When GPUs exchange data, the compute units that do the copying cannot run the model at the same time.

GPU kernels in RCCL and rocSHMEM can now hand transfers to dedicated SDMA copy engines. In AMD's tests on MI350X GPUs, bandwidth matched or beat the default method.

The quantized GLM-5.3 model needs about 464 GB of GPU memory, more than one GB300 GPU has.

A guide that links two DGX Stations over RDMA networking and splits the model between them with vLLM, serving it through one API endpoint.