🔥 Top Story
Run a vLLM Server on HF Jobs in One Command
Hugging Face shows how to launch a vLLM inference server on HF Jobs infrastructure with a single command. hf jobs run is effectively docker run for HF infrastructure: it pulls the official vllm/vllm-openai image, takes a GPU via --flavor, and exposes the port with --expose, yielding an OpenAI-compatible endpoint. There are no servers to provision and no Kubernetes to configure; billing is per-second and the endpoint is gated by your HF token rather than public (an a10g-large runs at 1.50 dollars per hour). Larger models scale by sharding across GPUs with --tensor-parallel-size — for example the 122B Qwen3.5 mixture-of-experts model on 2x H200. You can also SSH into a running job to debug, or point an agent harness like Pi at the endpoint to drive a coding agent on your own self-hosted model. The post frames HF Jobs as the choice for experiments, one-off evals, and batch generation, while Inference Endpoints, with scale-to-zero and finer-grained access control, fit durable production services.
💡 Why it matters: It lets teams stand up an OpenAI-compatible model endpoint on demand with per-second billing and no dedicated GPU infra or Kubernetes, cutting cost and lead time for evals, batch jobs, and prototyping.
🔗 Read more · Hugging Face
Kubernetes & Cloud Native
Building a Cluster-Aware AI Agent with Kubernetes, Argo CD, and GitOps
CNCF
CNCF Ambassador Maryam Tavakkoli published a hands-on guide to running a self-hosted, read-only AI agent inside a Kubernetes cluster. The runtime is an Ollama pod serving a local Mistral 7B model (port 11434), a FastAPI pod exposing the HTTP API and chat UI (port 8000), and a PersistentVolumeClaim for the model weights, with a dedicated ServiceAccount bound to a ClusterRole granting only get/list so the agent can observe but never mutate the cluster. It offers two modes: /ask answers from the LLM alone, while /diagnose is the true agent — it reads pod status, recent events, and logs from non-running pods in a namespace, injects that into the prompt, then reasons (RAG over live cluster state). Delivery is full GitOps: GitHub Actions builds a multi-arch image (amd64+arm64) tagged with the 7-character commit SHA, Argo CD Image Updater polls the registry every two minutes and commits new tags into kustomization.yaml, and Argo CD reconciles the cluster. Apart from the model pull at startup, there is no external egress, so no data leaves the cluster. The author urges starting every agent read-only — because hallucinations multiplied by write access invite incidents — and earning each additional permission one RBAC verb at a time.
💡 Constraining the agent with read-only RBAC and running the model inside the cluster bounds both data-exfiltration and misfire risk at the API-server level, making it a sound starting point for platform teams experimenting with AIOps.
AI & ML
Which tokens does a hybrid model predict better?
Hugging Face
Ai2 (the Allen Institute for AI) published a token-level study of which tokens a hybrid language model predicts better than a transformer. It compares Ai2's own 7B transformer (Olmo 3) and hybrid (Olmo Hybrid), deliberately matched in data, tokenizer, and training recipe so that differences mostly reflect architecture. A hybrid keeps a few attention layers but swaps the rest for recurrent layers, giving flat per-token cost regardless of input length, at the price of a compressed, lossy memory that cannot recall an exact earlier token the way attention can. The hybrid's advantage is largest on meaning-bearing tokens — nouns, verbs, adjectives — and on tokens requiring state tracking, like which entity a pronoun refers to (a loss gap around 0.04 for content words vs. about 0.02 for function words). Its edge nearly vanishes on tokens that verbatim-repeat earlier text and on closing braces, where attention (and thus the transformer) excels. The team proposes filtered, token-type-specific losses as a sharper way to compare architectures, with details in the tech report (arXiv:2606.20936).
💡 Architecture strengths and weaknesses that a single average loss hides show up by token type, so teams adopting hybrids should weigh the trade-off for copy- and exact-recall-heavy workloads like code and retrieval.
How agents are transforming work
OpenAI
OpenAI published a new economic research paper, 'The Shift to Agentic AI: Evidence from Codex,' measuring the economic potential of its coding agent Codex. Its central claim is that agents change the unit of knowledge work from short, self-contained chats to delegated, long-horizon tasks. By May 2026, 80.6% of sampled individual users made at least one Codex request estimated to exceed 30 minutes of human work, 70.2% exceeded one hour, and 25.6% exceeded eight hours. Inside OpenAI, every department — including Legal, Finance, and Recruiting, not just Engineering — crossed over to Codex as its primary AI tool around April 2026, and Codex now accounts for 99.8% of weekly output tokens generated within the company. Non-developer adoption grew especially fast, rising 137x for individual users and 189x for organizational users since August 2025. The heaviest users at the 99th percentile regularly generate more than 60 hours of Codex agent work per day across multiple parallel agents.
💡 A usage pattern where agents run multi-hour tasks in parallel explodes token consumption, CI load, and review bottlenecks, signaling that platform, FinOps, and code-review processes need redesigning for agent scale.
Cloud Updates
STOCKSTAY Another Day: The Latest Addition to Turla’s Intelligence Gathering Apparatus
Google Cloud
Google Threat Intelligence Group (GTIG) published an in-depth analysis of STOCKSTAY, a backdoor continually developed and deployed by the Russia-linked threat actor Turla. STOCKSTAY was originally designed to masquerade as a stock-market data viewing tool, baking that disguise into its file-naming scheme and into how it stores implant configuration, control messages, and response data. Early versions kept this stock theme internally, but in 2025 GTIG identified variants masquerading as other benign applications such as PDF viewers and calculator utilities. The malware is built from components like STOCKSTAY.STOCKBROKER and STOCKSTAY.STOCKMARKET, and its encrypted configuration file falsely describes itself as an application for trading information. GTIG frames STOCKSTAY as the latest addition to Turla's intelligence-gathering apparatus. The report includes the malware's architecture overview and decrypted configuration format.
💡 A state-linked backdoor that disguises itself as legitimate utilities and encrypts its config evades signature detection, so defenders need behavior-based detection and threat hunting using the IOCs and config formats GTIG disclosed.
DevOps & Infrastructure
Deploy Boundary on Kubernetes with official Helm charts
HashiCorp
HashiCorp released two official Helm charts for deploying and operating Boundary on Kubernetes. Until now, teams had to hand-craft and maintain separate manifests — deployments, services, config maps — for each component and keep them consistent over time. The new controller chart deploys Boundary controllers onto a cluster, while the worker chart deploys Boundary workers. Workers installed via the chart can connect to any Boundary controller, including HCP-managed controllers, self-managed controllers on VMs, or controllers deployed onto Kubernetes by the controller chart. The charts shipped alongside the Boundary 1.0 release. Because Boundary is designed as a distributed system, it can run on any runtime, including containers on Kubernetes.
💡 Replacing bespoke manifests with official charts standardizes deploying and upgrading Boundary on Kubernetes, and the worker chart's ability to attach to any controller simplifies hybrid PAM topologies.
“Code should be regenerated, not maintained”: Codeplain makes the case for spec-driven development
The New Stack
The New Stack profiles Codeplain, which argues that code should be regenerated, not maintained. As AI generates code faster than teams can review it, a small but growing group of developers argues for treating a specification as the single source of truth rather than hand-editing generated code. Codeplain is built on Plain, an open-source specification language that captures how software should behave in structured, human-readable documents. When something breaks or needs to change, you edit the spec — not the code — and Codeplain regenerates the implementation from scratch. The article illustrates this with a Plain spec for a Trello client and the Python code generated from it. It is positioned as one answer to the ongoing debate over who maintains AI-generated code.
💡 Treating the spec as the source of truth and regenerating implementations sidesteps the maintenance debt of AI-written code, but it shifts the risk onto spec quality and the determinism of regeneration.
Trust in Rust: Foundation debuts official training to tackle steep learning curve
The New Stack
The Rust Foundation announced the Rust Foundation Trusted Training (RFTT) program, a formal accreditation for Rust training providers backed by the nonprofit that stewards the language itself. It is positioned as an official quality signal for a Rust training market that has grown rapidly alongside the language's adoption. Executive Director and CEO Dr. Rebecca Rumbul says the program builds trust by ensuring that those who claim to be experts actually are. The foundation notes that Rust adoption has accelerated across industries — from systems programming and embedded development to safety-critical infrastructure — driving demand for training. RFTT aims to give learners, employers, and the broader community a trusted way to evaluate training quality. The move directly targets Rust's well-known steep learning curve as a barrier to adoption.
💡 As Rust spreads into safety-critical domains, an official accreditation gives teams a reference point for vetting skills in hiring and internal training, reducing adoption risk.
Template-based data extraction is dead. Here’s what comes next.
The New Stack
The New Stack argues template-based extraction is dead and introduces Amazon Bedrock Data Automation (BDA). BDA is a fully managed, generative-AI-powered AWS service that automates extraction, classification, and transformation of unstructured content across modalities — documents, images, audio, and video. Foundation models sit at its core to intelligently extract and understand content, and users can rely on standard output for common cases or define custom logic with business-specific blueprints. The piece frames the shift around how document formats have diversified — PDFs, contracts, scanned images, call recordings, meeting videos — making rigid, rule- and template-based pipelines brittle and costly. BDA is designed for scalability, accuracy, and auditability, which the author says makes it well suited to enterprise workflows. Standard output returns the model's default response directly from the Data Automation pipeline.
💡 Moving unstructured-data pipelines from brittle templates to a managed FM service cuts maintenance, but the real test becomes validating extraction accuracy, audit trails, and per-document cost against your actual document mix.
How we cut AI costs by 80%
The New Stack
In a New Stack contribution, a team describes cutting enterprise AI agent costs by roughly 80% using a context lake approach. The core argument is that rising AI costs come not just from model prices but from how each call fills the context window. Assembling many data sources on the fly and stuffing them into the prompt inflates token counts and cost. Using pre-joined data instead simplifies the context window and optimizes token usage. The result is handling the same work with far fewer tokens, for an roughly 80% cost reduction.
💡 It reframes agent cost optimization around the context-assembly step rather than model choice, suggesting that pre-joining data to shrink tokens can be an effective FinOps lever.
Agent Toolkit for AWS includes 20+ agent skills, but your agent might not load them without this one file
The New Stack
The New Stack highlights that the Agent Toolkit for AWS ships 20+ agent skills, but a single 17-line rules file determines whether your agent actually uses them. The toolkit was released on May 6, works with Claude Code, Codex, Kiro, and any MCP-supporting agent, and is organized into three layers. The author admits to skipping the README and jumping straight in, missing the very file that makes an agent reliably reach for the toolkit's tools. Without that rules file, the 20-plus skills may be installed yet go unused because the agent does not predictably invoke them. The takeaway is that shipping tools is not the same as getting an agent to use them. One small configuration file ends up deciding the reliability of the agent's behavior.
💡 Wiring tools into an agent is not enough — you must steer invocation via a rules/system-prompt file, an operational detail teams routinely miss when integrating MCP-based agents.
Boundary 1.0 releases RDP session recording and improved management
HashiCorp
HashiCorp announced the general availability of Boundary 1.0, its privileged access management (PAM) solution. Like other HashiCorp 1.0s, the milestone is less about a single feature than about maturity in user experience, use-case coverage, architecture, and proven production deployments. The headline new capability is RDP session recording, which captures and replays every interactive action in an RDP session to support compliance and deeper security analysis, with Kubernetes, database, and HTTP/HTTPS protocols planned next. The release also adds two Kubernetes Helm charts, scoped aliases for org and project scopes, and an enhanced admin UI that guides permission grants with dropdowns and reusable templates. Session recording first arrived for SSH in 0.13, and RDP credential injection landed in 0.22, so this extends a steady push into Windows environments. HashiCorp (now part of IBM) also previews a roadmap toward securing AI agents and non-human identities (NHIs) with HTTP credential injection and step-by-step, continuously re-evaluated authorization.
💡 RDP session recording and K8s Helm charts ease auditing and deployment in Windows and container environments, while the NHI/agent-access roadmap signals that PAM strategy needs rethinking as machine identities outnumber humans.
Scaling without friction: Aliases at project scope in Boundary
HashiCorp
Alongside Boundary 1.0, HashiCorp introduced aliases at org and project scope. Aliases let end users connect to targets using a memorable, DNS-like name instead of a random generated ID, but alias names previously had to be globally unique across Boundary. That meant different teams in a multi-tenant deployment could not reuse the same alias name. The new feature has Boundary automatically append scope-based suffixes, so aliases stay globally unique while teams in separate org or project scopes can use similar names — for example mylinuxhost.rnd.dc-canada and mylinuxhost.marketing.dc-canada. As a result, customers can connect to targets via aliases at scale without naming conflicts between teams. It is a quality-of-life improvement aimed at large, multi-team deployments.
💡 Resolving the global-uniqueness constraint with scope-based suffixes cuts alias naming collisions and operational friction in multi-tenant setups, smoothing self-service access for large teams.
HCP Vault Dedicated introduces cluster disaster recovery (public preview)
HashiCorp
HashiCorp added cluster-level disaster recovery (cluster DR) to its fully managed HCP Vault Dedicated service as a public preview. Where existing regional DR assumes large-scale infrastructure failures like cloud or network outages, cluster DR targets cases where the cluster itself is the problem, enabling cluster-level failover and DR drills. During the preview it is available for production-tier clusters with DR enabled; failover and failback are requested through HashiCorp support, and requests within 16 hours of an incident are routed to the on-call Vault team. It also supports security response — isolating a suspected-compromised cluster, promoting the DR secondary, and restoring continuity under controlled conditions. HashiCorp recommends combining regional DR, which protects where Vault runs, with cluster DR, which protects how Vault operates. The feature reflects the reality that most disruptions stem from software, operational, or security issues rather than full regional outages.
💡 Since Vault is the control plane for auth, dynamic secrets, and encryption, rehearsing cluster-level failover to validate runbooks and recovery steps reduces the risk that Vault downtime cascades into halted deployments and broken service connectivity.
This digest was collected from RSS feeds and summarized by AI (Claude). See the original links for full details.