Security teams confront "wild code" from employee-built AI agents

Datadog's Propolis uses agents to handle testing as AI-generated code volume surges; Anthropic engineers now ship eight times as much code as before, with some shipping thousands of PRs per month.

In a CNN "Terms of Service" interview, Microsoft AI CEO Mustafa Suleyman lays out his case for a "human-first" AI future as development accelerates. The conversation opens on bringing a child into the AI era and moves into facing AI's risks.

After two days of LoRA training, the user says Moss never really converges on likeness and has trouble with characters and fine details, though styles and concepts seem fine. Datasets previously used with Krea were tried.

OpenAI Alignment adds pair-matched comparisons and post-control risk to g-mean², its dataset-level measure of monitor discrimination. The post derives a pessimistic lower bound for intervention evaluations, where g-mean² previously used only the optimistic upper bound.

AWS's Agent Registry lets agents discover tools at runtime instead of hardcoding them, targeting enterprise systems that expose hundreds of internal APIs for pricing, support and orders. AWS says hardcoded tools degrade at a few dozen and become unmanageable beyond that.

Announced at PyTorch Conference 2026 and co-developed with Meta's PyTorch team, TorchTPU replaces PyTorch/XLA and targets TPUs via a standard torch.device("tpu"). Google reports 52% MFU training Qwen 3 8B with TorchTitan on Ironwood, scaled to 8,192 TPU7x chips.

Arm built open-source adapters for TOSA, VGF, and PTE formats, letting engineers inspect Arm-specific IRs and hardware profiling traces in Model Explorer's hierarchical graph view. Arm used it to catch unintended TRANSPOSE operators introduced by an ExecuTorch Arm backend change in an exported NSS model.