Cloudflare’s new WAF integration hooks live threat feeds into its edge network so security teams can auto-block malicious IPs before they hit your servers. It pulls from millions of indicators in real time, lets you filter by threat actor or industry, and claims near-zero latency. Only Cloudforce One subscribers get it today, but Cloudflare says performance won’t dip even under heavy lookup loads. Anthropic rolled out two versions of its latest model, Claude Fable 5 and Claude Mythos 5. Fable 5 handles coding, vision, long-memory, and scientific queries with a conservative safety layer that routes sensitive requests to an older Opus 4.8 engine. Mythos 5 lifts some of those safety rails for vetted cyberdefense and life-sciences partners. Meanwhile, HashiCorp’s Boundary product tackles AI infrastructure access by issuing unique identities, just-in-time Vault credentials, and session-level controls—no more hard-coded secrets or overprivileged bots. On the platform side, Microsoft Foundry positions itself as a one-stop shop for production AI. You pick, test, route and monitor models against your own cost and quality metrics, all in a unified, model-agnostic interface. And a joint effort from Mirantis and Logsight.ai demonstrated cross-border GPU pooling—Nvidia A100s in Quebec talking to AMD MI300Xs in Atlanta—controlled from Frankfurt via the open-source k0smos stack. They even spun GPUs up and down based on real-time electricity prices. Several community tools and experiments showed what’s next. MemPalace stores conversation memory as text inside a local “palace” structure, hitting 96.6% recall on LongMemEval without any cloud calls. whichllm scans your CPU, GPU and RAM, then ranks HuggingFace models by real benchmarks, not just parameter count. GitButler’s Grit project rebuilt Git in Rust with coding agents, passing 41,715 of 42,001 tests but racking up 45 billion tokens and requiring heavy human oversight. On Kubernetes, the Inference Extension steers LLM traffic by KV cache status and queue depth, cutting latency. And Cilium locked down its CI/CD pipeline by pinning dependencies, separating trusted workflows, and requiring reviews for any change.