🔥 Top Story
In-House LLM Serving at Netflix
The Netflix technology blog covers building in-house LLM serving, describing the approach of operating their own serving infrastructure rather than relying on external APIs.
💡 Why it matters: More organizations at scale are choosing self-hosted serving — if request volume and latency requirements have crossed a threshold, it is worth recomputing total cost against managed APIs.
🔗 Read more · Netflix
Kubernetes & Cloud Native
Flipkart and LitmusChaos at KubeCon + CloudNativeCon India 2026: A recap
CNCF
CNCF recaps Flipkart and LitmusChaos at KubeCon + CloudNativeCon India 2026, held June 18–19 in Mumbai. Flipkart won the CNCF End User Case Study Contest and delivered a keynote titled "From Afterthought to Practice: How Flipkart Built a Multi-tenant Chaos Platform on LitmusChaos." The centralized chaos platform Flipkart built addressed four customizations: a hybrid multi-tenancy architecture, a DaemonSet-based high-availability model, a Script Runner fault for dynamic targeting, and a hybrid VM chaos extension for non-Kubernetes workloads. The LitmusChaos project pavilion drew between one and two hundred visitors. Discussion topics covered resilience testing for AI workloads, the ChaosHub fault library, CI/CD automation, and the newly introduced LitmusChaos MCP integrating the Model Context Protocol.
💡 Extending a multi-tenant chaos platform to cover non-Kubernetes VM workloads is the direction worth noting for organizations with mixed infrastructure.
AI & ML
Fine-tune video and image models at scale with NVIDIA NeMo Automodel and 🤗 Diffusers
Hugging Face
NVIDIA describes fine-tuning video and image models at scale with NeMo Automodel and Hugging Face Diffusers. The core capability is pointing pretrained_model_name_or_path at any Diffusers model ID on the Hub and starting training with no checkpoint conversion required. Supported models include FLUX.1-dev and 2-dev (12B–32B text-to-image), Wan 2.1 and 2.2 (1.3B–27B text-to-video), HunyuanVideo 1.5 (13B) and Qwen-Image (20B). Parallelism options — FSDP2, tensor parallel, expert parallel, context parallel and pipeline parallel — are selectable via configuration. On benchmarks, full FLUX.1-dev training reaches 35.51 images per second on 8× H100 GPUs and Wan 2.1 (1.3B) reaches 8.50 clips per second, with LoRA variants significantly reducing memory. It supports both full fine-tuning and parameter-efficient LoRA adapters, using pre-encoded VAE latents and multiresolution bucketing. It is open source under Apache 2.0 via Docker container or pip installation, with integration documented in the Diffusers training guides.
💡 Removing the checkpoint conversion step eliminates a large share of practical friction — enough to pull forward plans to trial fine-tuning a Hub model.
A scorecard for the AI age
OpenAI
OpenAI CFO Sarah Friar lays out a scorecard for the AI age, opening from the question she hears from CFOs everywhere: how do we get more value from our AI spend? For years the market measured software success through adoption — seats purchased, users active, licenses renewed — but understanding the value of AI demands a more powerful measure: work accomplished. The basic economic question is whether the value of the work AI completes grows faster than the cost of producing it. Answering that requires looking deeper than a metric such as cost per token: a lower-cost model may have cheaper tokens but require more attempts, more time or more human review to get good results, while a more capable model may have more expensive tokens but complete the same task in one pass. What matters is the full cost of producing a successful outcome measured against the value that outcome creates. The scorecard consists of four questions: how much useful work gets done, what a successful task costs, whether people can depend on the result, and whether each AI dollar produces more value as usage grows. The recommended starting point is picking one workflow, defining what "done" means, and measuring that outcome in the system where the work actually happens.
💡 The observation that cheaper models can require more retries and human review is the practical core — rebuild model comparisons around total cost per successful task rather than per-token price.
Cloud Updates
Cloudflare WAF protects WordPress applications from two high-severity vulnerabilities
Cloudflare
Cloudflare deployed two WAF rules in response to high-severity vulnerabilities disclosed by the WordPress security team: CVE-2026-60137, a high-severity SQL injection, and CVE-2026-63030, a critical unauthenticated remote code execution. The SQL injection affects WordPress 6.8 and later while the RCE affects 6.9 and later; versions before 6.8 are unaffected. Patches shipped in WordPress 7.0.2 with backports to 6.9.5, 6.8.6 and 7.1 Beta 2. The WAF rule IDs are 1c060d3a371549219ee290d7ed933fcc and 7dfb2bd4708d4b88b9911dc0550664b6 in the Managed Ruleset, and db003b39b7774859a8d588ce33697a1a and ebd3f2df15c74ddcbf6220c9b5ec246a in the Free Ruleset. Protection covers all Cloudflare customers on affected WordPress versions, free and paid plans alike, whose traffic routes through the WAF. The rules were activated on July 17, 2026 at 17:03 UTC with a default Block action.
💡 With an unauthenticated RCE in the pair, sitting behind a WAF is no reason to defer patching — and since only pre-6.8 is unaffected, start from a version inventory.
Eclipse Dataspace Components on AWS: Cost optimization strategies
AWS Architecture
AWS covers cost optimization strategies for running Eclipse Dataspace Components (EDC) on AWS, opening from the challenge of predicting and controlling infrastructure cost when deploying EDC connectors. The baseline cost drivers are Amazon Aurora PostgreSQL at $276 per month for business-critical, Amazon ECS with Fargate at $83 per month, and a Network Load Balancer at $20 per month. For non-critical workloads, using smaller database instances (db.t4g.medium) and AWS Fargate Spot capacity yields up to 58% savings and reduces compute by 70%. Aurora Serverless v2 automatically scales database capacity to application needs, eliminating peak-capacity provisioning. Fargate Spot can save up to 70% for workloads that tolerate interruptions, particularly development and non-critical environments. For storage, S3 Lifecycle policies transition infrequent data to lower-cost classes such as Intelligent-Tiering and Glacier Instant Retrieval. Savings Plans for Aurora and Fargate provide significant discounts for predictable, steady-state business-critical connectors.
💡 The breakdown showing Aurora at $276/month as the dominant line item is the useful part — if you plan to run multiple dataspace connectors, settle the instance-tier policy first.
Eclipse Dataspace Components on AWS: Architecture patterns in production
AWS Architecture
AWS lays out architecture patterns for running Eclipse Dataspace Components connectors in production on AWS, premised on the need for deliberate decisions around isolation, managed services and security layering. EDC connectors deploy the control plane and data plane as containers, orchestrated through Amazon ECS with AWS Fargate for serverless scaling. AWS Secrets Manager, Amazon Aurora and Amazon Cognito provide secrets storage and OAuth 2.0 credential vending as managed services. Amazon S3 handles shared data, while API Gateway and a Network Load Balancer provide secure, isolated access through VPC links within private subnets. The isolation model treats all cloud resources belonging to a single EDC connector instance as an isolated architecture cell, with distinct security contexts for control and data planes. Defense principles named are network isolation, IAM least privilege, encryption at rest and in transit, and component decoupling that enables independent failure recovery. Production validation comes from the Prometheus-X Data Space Connector, using ECS Fargate, S3, Lambda and EventBridge for education sector use cases.
💡 Treating each connector instance as an isolated architecture cell is the load-bearing choice — it bounds both failure and compromise propagation at the instance boundary in cross-organizational sharing.
Eclipse Dataspace Components on AWS: Data sharing fundamentals
AWS Architecture
AWS opens a three-part series on implementing Eclipse Dataspace Components on AWS with data sharing fundamentals. EDC provides technical components implementing data spaces per International Data Spaces Association (IDSA) standards, enabling secure cross-organizational data sharing. The series covers foundational concepts in part 1, production AWS deployment patterns in part 2 and cost optimization in part 3. Core components are a federated catalog, connectors comprising control and data planes, and identity hubs managing participant credentials. The protocols are the Dataspace Protocol (DSP) for data exchange and the Decentralized Claims Protocol (DCP), which establishes trust using Decentralized Identifiers and Verifiable Credentials. In the connector architecture, control planes negotiate contracts while data planes transfer information between provider and consumer entities across legal boundaries. AWS integration incorporates extensions for Amazon S3, Secrets Manager and DynamoDB into custom EDC builds through a modular Gradle-based architecture.
💡 Separating contract negotiation (control plane) from data transfer (data plane) is what makes it possible to audit who agreed to what independently of the transfer path.
What’s new with Google Cloud
Google Cloud
The "What's new with Google Cloud" page appeared in the digest. It is a continuously updated hub collecting the latest Google Cloud news, announcements, resources, events and learning opportunities in one location. Because the page is updated on an ongoing basis, the individual items present at collection time differ from its later contents.
💡 Continuously updated hub pages surface under the same URL on every RSS pass — worth deciding whether a digest pipeline should filter this class of item out.
Level Up Your Column-level Security: Using IAM Data Governance Tags in BigQuery
Google Cloud
Google Cloud describes improving BigQuery column-level security with IAM data governance tags, noting that many customers have relied on policy tags as the go-to solution for column-level access control. Data governance tags replace policy tags, offering global scope and scalability against policy tags being regional-only. The workflow is to create tag keys with --purpose=DATA_GOVERNANCE in IAM Resource Manager, build hierarchical tag values up to five levels deep (for example PII > Financial > CreditCardNumber), attach tags to BigQuery columns via JSON schema or SQL ALTER TABLE, then define regional data policies referencing tag values to grant or deny access or apply masking. Advantages include global scope with regional enforcement across projects, automatic disaster recovery replication to secondary regions, decoupled governance letting you tag data before enforcing access policies, and hierarchical classification for granular control. Data policies are regional and must match the table region, and users need base-level table access such as roles/bigquery.dataViewer before data policies take effect as a second security layer.
💡 Decoupling tagging from enforcement is the practically important property — it lets you complete sensitive-data classification across the estate before switching access control on.
13 hands-on demos to build on Gemini Enterprise Agent Platform
Google Cloud
Google Cloud published 13 hands-on demos for building on the Gemini Enterprise Agent Platform, introduced earlier in the year as the place to build, scale, govern and optimize agents. Four build demos cover the ADK Foundation codelab for a basic conversational agent with CLI and web UI, an ambient expense agent using a graph-based workflow with human-in-the-loop, FastAPI and Pub/Sub, Model Context Protocol integration with BigQuery, file search and APIs, and Agent-to-UI (A2UI) rendering dynamic visual components in real time. Four scale demos cover a stateful data science agent deployed on Agent Runtime with Memory Bank persistence, long-running agents with durable state machines and checkpoint-resume, deploying the expense agent to Agent Runtime via the Agents CLI with Cloud Trace, Logging and BigQuery analytics, and a Cloud Run frontend dashboard with an OIDC-authenticated Pub/Sub pipeline. Governance demos cover secure agentic coding with TDD, a STRIDE threat model, Semgrep pre-commit hooks and PreToolUse gates, plus an Agent Gateway with mTLS, IAP, IAM and Model Armor content inspection. Optimize demos cover a quality flywheel with five-stage evaluation using AutoRaters, cross-language multi-agent work joining Python and Go agents over the Agent-to-Agent (A2A) protocol, and multi-framework orchestration linking an ADK control room with LangGraph and CrewAI via A2A.
💡 Notably the governance demos reach down to code-level gates — Semgrep hooks and PreToolUse — contrasting with approaches that handle agent security through runtime policy alone.
Why your AI agent framework isn't enough: 7 platform capabilities missing from production
Red Hat
Red Hat argues that an AI agent framework is not enough and names seven platform capabilities missing from production. First is cryptographic identity: verifiable proof of workload identity with scoped credentials. Second is execution sandboxing, providing hardware and application-level isolation for compromised workloads. Third is tool governance, where infrastructure-level policy determines which tools an agent can reach. Fourth is observability and tracing, capturing full execution traces of prompts, tool calls and results. Fifth is continuous evaluation, scoring production outputs against policy and ground truth. Sixth is safety enforcement, with guardrails intercepting output before it reaches customers or databases. Seventh is lifecycle management for deploying, updating, scaling and retiring agents across fleets. Red Hat products mentioned are Red Hat AI, Red Hat OpenShift AI, Red Hat Enterprise Linux and Red Hat Ansible Automation Platform.
💡 The useful distinction is that five of the seven live in the platform layer rather than the framework — grounds to plan framework selection and platform readiness as separate tracks.
Introducing Red Hat build of Karpenter
Red Hat
Red Hat introduced the Red Hat build of Karpenter, framed around infrastructure efficiency and compute cost control being a continuous effort. Karpenter is a Kubernetes-native node autoscaler that provisions right-sized compute just-in-time based on pending pod requirements and then consolidates underutilized nodes. The supported platform is Red Hat OpenShift Service on AWS (ROSA) with hosted control planes, requiring OpenShift 4.22 or later. It is enabled with the OpenShift 4.22 release and can be activated on existing clusters after upgrading. It automatically provisions optimal EC2 instances based on CPU, memory and scheduling constraints, running its controllers in the hosted control plane rather than on worker nodes. Cost benefits cited are automatic right-sizing, just-in-time scaling, Spot Instance utilization, prioritization of AWS commitments such as ODCRs and Capacity Blocks for ML, and reduced operational overhead. It coexists with the Cluster Autoscaler for gradual migration and supports kubelet configuration, TuneD profiles and FIPS, SOC 2 and FedRAMP compliance.
💡 Coexistence with the Cluster Autoscaler is the condition that matters most for migration planning — it allows a node-group-by-node-group cutover with measured comparison.
Friday Five — July 17, 2026
Red Hat
Red Hat's weekly roundup for July 17, 2026. First, InfoWorld covered Red Hat OpenShift 4.22, the latest version of its hybrid cloud application platform, focused on cutting cloud infrastructure costs, simplifying operations of virtualized workloads and securing sensitive data. Second, on the Technically Speaking podcast Red Hat CTO Chris Wright and Jered Floyd discussed sovereign AI, examining four pillars of sovereign AI and addressing data ownership and platform autonomy. Third, Red Hat is partnering with the CDC and NIH to enhance the Model Context Protocol for enterprise use in public health infrastructure, aiming to make agentic AI systems secure, standardized and reliable. Fourth, Swiss IT managed services provider Centris accelerated delivery using Red Hat Developer Hub, Red Hat OpenShift, Red Hat OpenShift Pipelines and Red Hat OpenShift GitOps. Fifth, on Bloomberg Intelligence Tech Disruptors, Chris Wright and Accenture's Jefferson Wang discussed photonics-based networks and their potential to unlock the full potential of the AI economy.
💡 The item on hardening MCP for enterprise use with public health agencies stands out — a sign the protocol is starting to absorb regulated-industry requirements.
DevOps & Infrastructure
The cost of saying yes has changed
GitHub
GitHub's engineering blog argues that the cost of saying yes has changed: the cost of writing code dropped while the cost of owning it did not, and offers a framework for deciding which changes are actually cheap in the AI era. It sets out three categories — cheap to own (adding display fields, refactoring tested helpers), not cheap to own (authorization changes, data-retention semantics, anything touching privacy, billing or compliance), and the skill of pricing uncertainty, meaning knowing when to try, when to debate and when to reject. As concrete examples, surfacing an existing last_active_at timestamp on a settings page is typically cheap, while changing authorization behavior is never cheap. Generated patches function as price checks, revealing true implementation scope through the diff. The key shift stated is that the most expensive part of a small feature request used to be writing the code, and now it is usually the meeting about whether to write it.
💡 Treating a generated patch as a price check is the practical move — produce the diff first to reveal real scope, before spending the meeting on whether to do it.
Kimi K3 tops Arena’s coding leaderboard — and it’s open-weight
The New Stack
Moonshot AI unveiled Kimi K3 this week, and within hours of its Thursday debut the open-weight model climbed to the top of Arena's frontend coding leaderboard, outperforming leading closed-source systems in blind evaluations. Developers building with AI have largely relied on proprietary models such as Anthropic's Fable and OpenAI's GPT-5.6 Sol for their most demanding coding tasks, and this release suggests open-weight models may be catching up faster than expected. In Arena's blind evaluations K3 ranked ahead of Anthropic's Opus 4.8 and OpenAI's GPT-5.6 Sol on frontend coding tasks, and on the general text leaderboard it finished above Opus 4.8 and roughly even with Sol. It is an impressive debut but still just one benchmark, and the real test will come when developers start throwing production problems at it. Most AI coding tools can already connect to multiple models, but the toughest jobs still tend to end up on OpenAI or Anthropic models — if Kimi K3 matches that level, engineering teams gain another option to run in their own environment instead of sending every request to a third-party API.
💡 An open-weight model beating top closed systems in blind evaluation means self-hosting may no longer be a performance compromise — a genuine added option for organizations constrained on data egress.
1Password’s new browser integration for Claude changes how AI uses your credentials
The New Stack
1Password shipped a browser integration for Claude. The context is that as AI agents take on more tasks such as managing online accounts, authentication has become a practical engineering problem, with companies like Coinbase already running over a thousand agents in production. The feature lets a fleet of agents sign in without ever exposing passwords to the LLM — what the company calls a zero-exposure security framework. Credentials are decrypted only when needed and passed directly to the browser, so Claude completes an action without ever seeing the underlying password or one-time code. 1Password CTO Nancy Wang says decryption and filling happen locally on the user's Mac using 1Password's standard autofill engine: when an agent needs to authenticate, 1Password decrypts the credential on-device and injects it directly into the target website through a secure channel, and 1Password does not return the plaintext credential to Claude or place it in the model's context. When 1Password recognizes a compatible AI agent controlling the browser, the extension enters a restricted state called Agentic Mode to enforce least privilege, preventing the agent from browsing or searching the vault and limiting access to credentials explicitly granted for the current task. Claude access is granted per task, with users approving or denying each session through a single biometric prompt via Touch ID or a password.
💡 The design guarantee that plaintext credentials never enter the model context is the key property — a reasonable minimum bar to demand when granting agents sign-in capability.
“They’re dead if they don’t offer this”: DoorDash’s CLI for agents may be out of necessity
The New Stack
DoorDash is rolling out a terminal tool that lets agents place real orders on its platform. The context is that as AI agents evolve beyond writing code toward handling everyday tasks for people, the infrastructure to keep them on track must keep pace. Co-founder and CTO Andy Fang announced the command-line interface, dubbed "dd-cli," on X this week, available to macOS developers in the US and Canada via a waitlist. The CLI gives developers a direct way to wire DoorDash ordering into whatever agent they are building — a personal assistant could use dd-cli to search nearby restaurants, compare prices, add items to a cart and pay, all with a single command. DoorDash already lets AI help people shop through a Claude connector, a ChatGPT integration and its own in-app "Ask DoorDash" assistant, but in each case a person stays in the loop, reviewing an order before it goes through. The CLI removes that step, setting DoorDash up to serve not just human customers navigating an app but autonomous software acting entirely on someone else's behalf. While developers welcome agents that can act without per-purchase approval, the move sparked debate about whether it is actually good for DoorDash's bottom line, with Gergely Orosz commenting on X that he was surprised by the direction given the economics of food delivery.
💡 Deliberately removing the human order-review step is the crux — when designing agent automation that touches payment, where you leave an approval point is the risk design.
How we brought agentic workflows to Cloud SIEM with the Datadog MCP Server
Datadog
Datadog describes bringing agentic workflows to Cloud SIEM with the Datadog MCP Server. The MCP tools built cover detection rule authoring, signal investigation, and bulk triage actions that assign, close or archive up to 500 signals at once. Usage analysis showed 44% of messages involved rule authoring or editing, 25% of customers performed bulk triage, and 14% pushed against the 50-signal UI limit. For progressive disclosure, a schema tool filters detection rule schemas by product type and detection method, reducing token usage 41–47% versus serving the full catalog. The bulk triage tool was redesigned to accept search queries instead of signal ID lists, keeping IDs out of agent context and enabling 500-signal operations. A custom eval framework integrated with CI flags regressions beyond 5% of baseline and requires minimum eval coverage for new tools before production, with a baseline eval success rate targeted around 80%. The bulk triage redesign cut update time from over 2 minutes for 50 signals to roughly 1 minute for hundreds, and one customer triaged more than 350 signals in 9 minutes.
💡 Redesigning the tool to take a query instead of an ID list, keeping IDs out of agent context, is the key move — it shows that when building MCP tools, context budget is the ceiling on batch size.
This digest was collected from RSS feeds and summarized by AI (Claude). See the original links for full details.