🔥 Top Story
Opus 5 costs a third of the price — and that’s actually the problem
Anthropic launched Opus 5 on Friday, arriving two months after Opus 4.8 and following the June releases of Mythos 5, Fable 5 and Sonnet 5 to round out the new generation, leaving only the lightweight Haiku awaiting an upgrade. Smaller than the flagship Fable 5, Opus 5 is significantly cheaper and noticeably less restrictive, designed to work on programming tasks for much longer without constant human input. Priced at $5 per million input tokens and $25 per million output tokens, it upends the cost-to-performance ratio for agentic tasks. On coding and knowledge work evaluations such as Frontier-Bench and GDPval-AA it establishes a new state of the art, and on the OSWorld 2.0 computer use benchmark it outperforms every other model at any given cost, surpassing Fable 5's best result at just over a third of the price. On ARC-AGI 3, which requires solving novel problems, its score is three times the next best model's. Because it is less expensive to run, teams can afford to let it work through larger coding tasks — which, the article notes, means thinking differently about security. Anthropic says the model is much stronger at verifying its work and iterating carefully until it succeeds: during benchmark testing it was given an incomplete prompt and intentionally prevented from viewing a drawing of a machine part, and rather than giving up it wrote its own computer vision pipeline to reconstruct the part from the image data.
💡 Why it matters: Lower prices lead to handing agents larger tasks, which widens their execution radius — the right order is preparing isolation before banking the cost savings.
🔗 Read more · The New Stack
Kubernetes & Cloud Native
Agentic AI Needs Guardrails, Not Guesswork
Docker
Docker recaps a roundtable that brought enterprise security leaders together on agentic AI's biggest challenge: how to govern AI agents without slowing developers down. Participants were Docker CISO Mark Lechner, Warp founder and CEO Zach Lloyd, NanoCo co-founder and CEO Gavriel Cohen, and CISO community founder Moriah Hara. Moriah Hara said the business wants AI agents everywhere while CISOs are left tolerating some tools and praying nothing breaks. Zach Lloyd advocated cloud-based agent infrastructure with centralized visibility and access controls. Gavriel Cohen recommended a training-wheels approach starting with unpermissioned data, plus a minimum seven-day release age for images. Named Docker products are Docker Sandboxes (isolated, disposable MicroVM-based environments for agent execution), Docker AI Governance for agent governance across teams, Docker Hardened Images as secure base images, and the Docker MCP Gateway as a chokepoint for authenticating, authorizing and logging tool calls.
💡 Requiring a minimum seven-day release age for images follows the same logic as Dependabot's three-day cooldown — supply chain defense is converging on simply not adopting immediately.
My LFX mentorship journey with kgateway
CNCF
A CNCF blog post recounts an LFX mentorship journey with kgateway, written by an engineer for whom open source has been a defining part of a long career in the cloud native ecosystem. kgateway is built on Envoy Proxy and the Kubernetes Gateway API, managing traffic in Kubernetes environments on a powerful and extensible foundation. The mentorship scope was adding chaos engineering support through HTTP fault injection, enabling platform teams to deliberately inject failures such as delays and aborts into their traffic. Contributions included implementing fault injection via a TrafficPolicy extension covering delay injection, abort injection with HTTP and gRPC status codes, response rate limiting and per-route overrides. The mentee also completed BackendConfigPolicy merge semantics work and added configurable filter stage positioning for ExtProc.
💡 Injecting faults at the gateway layer allows resilience testing without touching applications, lowering the barrier to running chaos experiments.
OpenTelemetry has graduated… Now what?
CNCF
CNCF covers what comes after OpenTelemetry's graduation. OpenTelemetry achieved CNCF graduated status in May 2026, joining projects such as Kubernetes and Prometheus as a production-ready, enterprise-grade standard. Project statistics show over 12,000 contributions from more than 2,800 companies with hundreds of maintainers, ranking as the second-highest velocity CNCF project behind Kubernetes. Graduation requirements met include production adoption across major organizations, a documented governance model, established community health processes, completed security audits, stable versioned APIs, comprehensive documentation and a successful TOC review. The roadmap covers observability for agentic workflows via generative AI semantic conventions, expansion into browser and mobile observability, and enterprise-scale tooling through Weaver for telemetry schema governance and the OpenTelemetry Injector for zero-code instrumentation. Ecosystem growth included adding profiling as a signal, expanding the OTel Collector and Demo, and introducing OpAMP, the OTel Operator, Weaver and Arrow.
💡 Generative AI semantic conventions landing on the roadmap matters practically — if you are inventing your own agent instrumentation schema, it is grounds to weigh waiting for the standard.
Launch of the AI Infra SIG under the CNCF Japan chapter: First meetup and call for speakers
CNCF
An AI Infra SIG launched under the CNCF Japan chapter, announcing its first meetup and a call for speakers, prompted by AI advancing from generative AI to agents and driving growing demand for scalable, efficient infrastructure. The SIG focuses on cloud native AI infrastructure best practices, covering scheduling (DRA, Kueue), orchestration (JobSet, KubeRay), deployment platforms (KServe, llm-d), networking and agent infrastructure. The first meetup is October 1, 2026 from 18:00 to 21:00 JST at Garden Terrace Kioicho in Tokyo, with a livestream and free registration. The speaker call closes August 28, 2026, taking general sessions of 20–30 minutes and lightning talks of 5–10 minutes. Six co-leads from LY Corporation, IBM Research Tokyo, CyberAgent and Preferred Networks established the SIG under Cloud Native Community Japan. Its mission is to share optimization techniques, operational best practices and real-world experiences for AI infrastructure while strengthening upstream open-source contributions.
💡 The SIG scope spanning scheduling through agent infrastructure shows AI infrastructure discussion widening past GPU allocation into workload orchestration generally.
Cloud Updates
BGP ORIGIN attribute manipulation and its impact on the Internet
Cloudflare
Cloudflare investigated BGP ORIGIN attribute manipulation and its impact on the Internet. In-depth testing found that nearly 70% of BGP paths experience ORIGIN attribute rewrites by transit providers seeking traffic advantages. Across public BGP collectors, ORIGIN values break down as 89.8% IGP, 3.5% EGP and 6.7% INCOMPLETE. Direct peer analysis showed roughly 10% of peers manipulate ORIGIN to IGP, and 26% of the top 50 ASes engage in the practice. ORIGIN rewriting redirects traffic toward large Tier-1 ISPs, with IPv6 paths gaining 40% more routes through manipulating networks. RFC 4271 states the ORIGIN attribute should not be changed by any other speaker, yet the violation is widespread. Cloudflare concludes there is no valid technical justification for ORIGIN rewriting in modern Internet routing. Its recommendation is to require all BGP implementations to set ORIGIN as IGP on all received and advertised routes, and it proposes reviving the "Scrubbing BGP ORIGIN Attribute" draft to deprecate ORIGIN's influence on path selection.
💡 If 70% of paths are rewritten, ORIGIN has effectively lost its meaning as a signal — traffic engineering policies that rely on it need their premises rechecked.
Open Knowledge format v0.2 tackles agentic trust
Google Cloud
Google Cloud released Open Knowledge Format (OKF) v0.2. When OKF was introduced in June 2026, the premise was that it should carry the context agents need — table schemas, metric definitions, runbooks. OKF is a markdown plus YAML frontmatter specification for sharing structured metadata in a format-agnostic way so AI agents can read and write knowledge bundles. v0.2 adds five trust signal families letting agents judge trustworthiness: provenance for sources, trust for generated versus verified, freshness expressed via stale_after, lifecycle for status, and attestation through an Attested Computation type. The concrete signals are generated, verified, sources, status, stale_after, usage_count, last_modified, executor, attester and receipt. The headline feature, Attested Computation, carries a sanctioned computation plus a deterministic attester to verify that the exact SQL or query ran as declared, validated by a receipt containing job_id, executed_sql and result. v0.2 is additive, so v0.1 bundles work unchanged.
💡 Encoding freshness and verification state directly into agent-readable metadata is a structural fix for agents acting on stale runbooks.
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 gathering the latest Google Cloud news, announcements, resources, events and learning opportunities in one place. Because it updates continuously, the items present at collection time differ from its later contents.
💡 The same evergreen hub URL recurring across multiple days of the digest is grounds for deciding whether the collection pipeline needs a dedupe rule.
Updated Cyber Threat Actor Naming System
Google Cloud
Google Threat Intelligence Group (GTIG) began rolling out a unified naming schema for tracking threat actors, based on two-word cryptonyms. The first word is a unique, memorable term for the specific actor, preferably drawn from prior public reporting and randomly generated where none exists. The second word categorizes the threat cluster by origin or motivation: CASTLE for China, ION for Iran, NEPTUNE for North Korea, RELIC for Russia and COMET for cybercriminal groups. Examples include FIN11 becoming RAZOR COMET, APT28 becoming LAKE RELIC, APT33 becoming BLEAK ION and APT37 becoming PLAIN NEPTUNE. The rollout prioritizes renaming several dozen of the most active groups on a rolling basis, with previous names remaining searchable in the GTI platform and MITRE ATT&CK mappings preserved. The "UNC" (uncategorized) designation continues for threat clusters in early investigation stages. It was announced July 25, 2026, with a July 30 update appending a comprehensive naming table.
💡 Old names remaining searchable is the practical detail — existing threat intel documents and detection rules need not be rewritten immediately, though mixed naming becomes something to manage.
Gain stronger pod isolation on Microsoft Azure Red Hat OpenShift with OpenShift sandboxed containers
Red Hat
Red Hat describes gaining stronger pod isolation on Microsoft Azure Red Hat OpenShift with OpenShift sandboxed containers, starting from the premise that running containerized workloads on OpenShift already provides industry-leading process isolation. The feature runs each pod inside its own lightweight VM with a dedicated kernel, layering hypervisor-backed workload isolation on top of existing OpenShift security controls. The technology foundation is Kata Containers, the open source project that runs each pod inside its own lightweight VM. It is available on Azure Red Hat OpenShift, with the OpenShift sandboxed containers operator handling installation, configuration and lifecycle management. Deployment means installing the operator from OperatorHub, creating a KataConfig resource and adding runtimeClassName to pod specifications. Recommended use cases are untrusted code, privileged containers, multi-tenant workloads, and AI or agent workloads that generate unreviewed code at runtime. Sandboxed pods cannot consume more CPU or memory than allocated to their VM and can only access explicitly passed-through devices.
💡 Naming agent workloads that generate unreviewed code at runtime as a recommended use case supports drawing the agent isolation boundary at the VM rather than container level.
Friday Five — July 24, 2026
Red Hat
Red Hat's weekly roundup for July 24, 2026. First, registration opened for the Red Hat OpenShift Commons Gathering Salt Lake City 2026, held alongside KubeCon + CloudNativeCon North America, bringing together the global OpenShift community including users, contributors, partners and Red Hat experts to share knowledge and production experiences. Second, Red Hat announced five partner program enhancements designed to streamline daily operations, safeguard investments and give partners a clear financial runway. Third, Red Hat's Vincent Danen discussed in an interview how AI agents reshape enterprise risk, noting AI is already embedded in customer support, software development and operational workflows with evolving security requirements. Fourth, a piece on quantum computing preparedness frames cryptographically relevant quantum computers as an urgent timeline security teams must plan for today, warning that bad actors are already capturing and storing encrypted enterprise traffic. Fifth, the Kubernetes Fleet Management Report 2026 examines organizational management across complex distributed environments, noting the challenge has shifted from adoption to operations — running clusters consistently at scale.
💡 With the harvest-now-decrypt-later threat model spelled out, there is little basis left for deferring a post-quantum migration timeline for data needing long-term confidentiality.
Beyond the blind spots: Defeating frontier AI model threats in your application development process
Red Hat
Red Hat addresses threats frontier AI models pose to the application development process, opening from how much has changed in cybersecurity in just a few months. The threats named are frontier models (Claude Mythos) autonomously discovering vulnerabilities in open source codebases at machine speed; vulnerability reporting compressed from a year into two-week windows, doubling annual triage volume; DIY Kubernetes fragmentation creating a fog of war across hundreds of thousands of nested CVE dependencies; and lateral movement exploiting wide-open east-west container traffic. Recommended defenses are shifting security left through developer pipeline integration, microsegmentation with zero-trust network policies, enforcing cryptographic signing of artifacts and container images, and using pre-hardened golden path templates. Red Hat products named are OpenShift with automated patching, kernel security fixes and immutable CoreOS; Lightwell for AI-driven remediation with backported patches; Trusted Artifact Signer for code signing; Trusted Profile Analyzer for SBOM auditing; Advanced Cluster Security for runtime monitoring; Developer Hub for compliance templates; and Quay as a centralized component repository.
💡 The observation that vulnerability reporting compressed from a year into two weeks and doubled triage volume is the core point — manual CVE triage cannot keep pace with that rate.
DevOps & Infrastructure
Jensen Huang made his first X post. He used it to lobby Washington about open-weight AI.
The New Stack
NVIDIA CEO Jensen Huang used his first post on X to share a public letter backing frontier open-weight models. Signed by Microsoft, Meta, Hugging Face and 22 other organizations, the letter argues that open models improve security, encourage faster innovation, and give countries and enterprises more control over their AI infrastructure. The Friday timing is notable as more organizations choose to run open-weight models inside their own environments rather than relying entirely on hosted services. The letter compares open-weight AI to open-source software, arguing that broad access drove decades of software innovation and the same principle applies to AI: instead of sending every request to a hosted model, organizations can download an open-weight model, run it on their own infrastructure, customize it for a specific workload and keep sensitive data behind their own firewall. NVIDIA has been moving in this direction, with its leaders describing a future where local and frontier models split work by cost, speed and control. As the coalition puts it, open weights let every organization match the right model to the right job at the right cost, reserving frontier-scale capability for genuine frontier problems and running efficient, specialized models everywhere else.
💡 Reserving frontier-scale capability for genuine frontier problems is a usable allocation rule — if every request currently goes to the top model, start by classifying work by difficulty.
Anthropic’s Opus 5 is almost Fable 5
The New Stack
Anthropic launched Opus 5 on Friday, the latest version of what used to be the company's flagship model before Fable 5. Anthropic says Opus 5 comes close to Fable 5's performance in many domains at half the price, with token costs unchanged from Opus 4.8 at $5 and $25 per million input and output tokens. Unlike Fable 5, users do not have to opt into a 30-day data retention policy to use it, and it is now the default model for Claude Max subscribers and the best model Claude Pro subscribers can access. Anthropic stresses it is the most capable Opus yet and can autonomously perform work for much longer than before, checking its own work and recovering from errors. Fable 5 remains the model for the most ambitious work and projects needing days-long autonomy, while Opus 5 is described as designed to be used every day. On virtually every benchmark Anthropic has shared, Opus 5 actually outperforms Fable 5 — suggesting the distinguishing factor is the ability to stay on one project for a very long time. It scores 1861 on the GDPval-AA v2 knowledge work benchmark, ahead of Fable 5 (1747), GPT-5.6 Sol (1736) and Opus 4.8 (1593). It is also the top model on Zapier's AutomationBench, which measures agents on end-to-end business workflows, where Anthropic says its pass rate is double the next-best model's at the same cost per task.
💡 A lower-tier model outscoring the flagship while the tiering persists is the signal — select models by sustainable autonomous working time rather than by benchmark score.
This digest was collected from RSS feeds and summarized by AI (Claude). See the original links for full details.