You're reading the Wednesday, November 18, 2026 edition. Today's briefing →
Toronto
Wednesday, November 18, 2026No. 157

Digital Plumber

Plumbing the information age

AI-curated intelligence for people who run networks. Daily coverage of AIOps, network automation, agentic operations, AI infrastructure, security and the vendors shaping them.

Today's 3 things that matter

Picked by the AI editor
  1. Routing·Analysis

    Microsoft 365 outage lasted 24 hours after gateway failover defect

    A major global Microsoft outage spanning November 12-14, 2026 lasted 23 hours 50 minutes, affecting Exchange Online, Teams, and downstream services.

    Why it matters Understanding interactions between automatic failover logic, shared infrastructure, latent defects and partial capacity is critical for SRE teams designing multi-service systems. Single points of failure in shared gateway infrastructure require robust isolation and staged rollout controls.

  2. Startups·Primary source

    OpenObserve raises $10M and adds autonomous AI SRE

    OpenObserve announced a $10 million Series A led by Nexus Venture Partners and Dell Technologies Capital, alongside Observability 3.0 capabilities including an autonomous AI SRE, anomaly detection, and LLM observability.

    Why it matters AI-native observability with autonomous AI SRE eliminates manual signal-from-noise sorting for teams managing petabyte-scale telemetry and LLM workloads.

  3. Research·Research

    Most LLM ops agents remain copilots rather than autonomous operators

    A new survey examines the transition from AIOps pipelines to agentic service operations through three coupled lenses: capability, operational authority, and evaluation evidence.

    Why it matters AIOps and SRE leaders evaluating autonomous remediation claims need this framework to assess actual decision-making authority and evidence quality before deploying LLM-based operations tools in production.

Today's briefing

What happened, and why it matters

15 stories · 9 topics · Updated 1:55 PM ET

Oracle's OCCL plugin lifts NCCL collective bandwidth up to 14%

SDxCentral · Nov 11, 2026 · Industry news

Oracle released Oracle Cloud Infrastructure Collective Communications Library (OCCL), a network plugin that improved NCCL AllReduce bandwidth by up to 10% and AllToAll by 14% on 16–128 node GB300 clusters, deployed atop OCI's multiplanar RDMA architecture with in-transport congestion control and NIC failure recovery.

Why it matters OCCL provides production-tested collective communications offload for OCI GPU clusters at scale, reducing network-induced delays in multi-node all-reduce patterns; multiplanar redundancy and transport-layer congestion control narrow the performance gap between OCI Spectrum-X Ethernet and proprietary NVIDIA Quantum InfiniBand on NCCL workloads.

Oracle announced Oracle Acceleron enhancements and OCCL at Oracle AI World 2026 (Nov 4), releasing a network plugin designed to run on OCI's multiplanar RDMA infrastructure. OCCL provides application-level congestion control and failure recovery at the collective-communications transport layer, enabling frameworks to continue using standard NCCL or RCCL APIs while OCI's plugin handles multi-plane network intelligence. Benchmarks using NVIDIA GB300 systems across 16–128 nodes showed 10% bandwidth gains for AllReduce (parameter synchronization in distributed training) and 14% for AllToAll (gradient exchange). OCI deployed 850 MW of data center capacity and 300,000+ GPUs in fiscal Q1 2027. For practitioners on OCI, OCCL now provides multi-plane redundancy with in-transport congestion management on collective operations at scale, reducing single-plane saturation from 100-node all-reduces and simplifying tuning of lossless Ethernet (PFC, ECN) across multiple independent network planes.

Read the original at sdxcentral.com ↗

Broadcom pushes Scale-Up Ethernet Transport for multi-vendor AI fabrics

HPCwire · Nov 11, 2026 · Industry news

Broadcom hosted SC26 workshop coverage of open AI and HPC fabric standards, emphasizing Scale-Up Ethernet Transport (SUE-T) under Open Compute Project and Ultra Ethernet Consortium specifications for deterministic, lossless transport, load balancing, congestion control, resilience, and telemetry across multi-vendor environments.

Why it matters SUE-T standardization under Open Compute Project and UEC 1.0 certification enable practitioners to architect multi-vendor AI fabrics with deterministic transport guarantees and certified interoperability; vendors can now ship UEC-compliant fabric switches, NICs, and telemetry stacks without proprietary lock-in.

Broadcom's SC26 workshop (Nov 15–20, 2026) emphasized standards-based approaches to AI and HPC fabric design, covering Scale-Up Ethernet Transport (SUE-T) under Open Compute Project, Ethernet for Scale-Up Networking (ESUN), Ultra Ethernet specifications, transport protocols, telemetry, and fabric management tools. SUE-T complements UEC 1.0 as a collective-communications and transport-layer standard, providing deterministic, lossless transmission guarantees. The implications are substantial: practitioners can now specify fabric using OCP SUE-T stack (open-source transport implementations) with UEC-certified NICs (Thor Ultra, Marvell, Pensando) and Broadcom Tomahawk 6 switches, avoiding NVIDIA or InfiniBand proprietary dependence. By late 2027, the Open Compute Project and UEC certification ecosystem will enable vendor-neutral fabric procurement with multi-vendor interoperability validation, simplifying procurement and reducing optics and switching costs through competition.

Read the original at hpcwire.com ↗

Quad9 pulls DNS resolver out of France over blocking orders

Cloudflare Blog · Nov 17, 2026 · Primary source

Quad9 announced it will withdraw its free public DNS resolver from France on November 24, 2026, rather than comply with court orders to block domains for copyright infringement. A French court imposed daily financial penalties, forcing the Swiss non-profit to choose between penalties or service changes that would harm privacy and security.

Why it matters Belgian courts have ordered Cloudflare, Google and Cisco to implement DNS blocking by January 2027 or face penalties. DNS operators globally must plan for multi-provider redundancy across jurisdictions as regulatory pressure on public resolvers intensifies.

Quad9's planned exit from France on November 24, 2026 sets a precedent for public DNS operators facing regulatory pressure across Europe. Cisco previously withdrew OpenDNS from France and Portugal and paused service in Belgium after similar legal pressure. As a small non-profit offering only a public resolver, Quad9 had no alternative compliance methods. The implications extend globally: a Belgian court has ordered Cloudflare, Google and Cisco to implement DNS blocking by early January 2027 or face penalties. For practitioners, this underscores the need for resolver redundancy across providers and jurisdictions—no single provider's service can be relied upon as infrastructure in regions with increasingly aggressive content-filtering mandates. Cloudflare, Google, Cisco, and Quad9 have all opposed mandatory blocking, with some earlier success: courts in Germany ruled against DNS blocking requirements in favour of Cloudflare and Quad9.

Read the original at blog.cloudflare.com ↗

Microsoft 365 outage lasted 24 hours after gateway failover defect

Office365itpros · Nov 15, 2026 · Analysis

A major global Microsoft outage spanning November 12-14, 2026 lasted 23 hours 50 minutes, affecting Exchange Online, Teams, and downstream services. Three conditions combined: Teams automatically failed over to the primary gateway (also serving Exchange Online) ~15 hours before the incident, a latent defect in gateway code (deployed July 2026), and 14% of capacity already offline.

Why it matters Understanding interactions between automatic failover logic, shared infrastructure, latent defects and partial capacity is critical for SRE teams designing multi-service systems. Single points of failure in shared gateway infrastructure require robust isolation and staged rollout controls.

Microsoft's preliminary Post-Incident Report revealed three converging factors: Teams failed over from its alternative gateway to the primary gateway—which also serves Exchange Online—about 15 hours before the incident; a defect in gateway code began rolling out in July 2026 and spread globally over weeks; and approximately 14% of gateway capacity was already offline when the cascade began. The outage affected businesses, airlines, and government agencies; email recovery took 16 hours before queues approached normal levels by November 13 afternoon; full restoration occurred by morning of November 14. This incident illustrates the risks of shared gateway infrastructure during automatic failovers when a dormant defect and partial capacity constraints intersect. Practitioners should carefully review failover policies and ensure defect detection during gradual rollouts.

Read the original at office365itpros.com ↗

Browsers distrust eMudhra after intermediate CA keys found guessable

Microsoft Security Response Center / Google Chrome Root Program · Nov 16, 2026 · Primary source

Microsoft is removing the remaining eMudhra root certificates across its products and services starting November 16, 2026. Eight eMudhra intermediate CA private keys—six for TLS, two for code signing—were compromised with trivially guessable values below 256, affecting approximately 30,800 live certificates primarily in higher education.

Why it matters Compromised TLS intermediate CAs enable MITM attacks on encrypted channels; audit certificate chains for eMudhra roots, implement OCSP stapling monitoring, and ensure trust stores are updated before Nov 16 to prevent acceptance of forged certificates.

Eight eMudhra intermediate CA keys were recovered and are now subject to mass distrust across browser root programs and OS trust stores. The key generation weakness exposed approximately 30,800 live certificates primarily in higher education and enterprise sectors. Forged certificates could intercept encrypted traffic, escalate trust across federated identity systems, move laterally across institutional networks relying on InCommon-issued certificates, and establish covert command-and-control channels behind trusted TLS. All major browser vendors, Microsoft, and language runtime maintainers scheduled or executed distrust implementations between November 12–18, 2026. Infrastructure teams must audit certificate chains for affected roots, automate renewal via ACME/ARI, and test failover to alternative CAs immediately. This represents a systemic PKI trust failure requiring rapid inventory and remediation across payment processors, e-commerce platforms, and secure communication channels.

Read the original at microsoft.com ↗

OpenObserve raises $10M and adds autonomous AI SRE

OpenObserve Blog · Nov 19, 2026 · Primary source

OpenObserve announced a $10 million Series A led by Nexus Venture Partners and Dell Technologies Capital, alongside Observability 3.0 capabilities including an autonomous AI SRE, anomaly detection, and LLM observability. Over 7,000 organizations now run the platform, including companies ingesting petabytes of data daily.

Why it matters AI-native observability with autonomous AI SRE eliminates manual signal-from-noise sorting for teams managing petabyte-scale telemetry and LLM workloads.

OpenObserve's $10M Series A from Nexus Venture Partners and Dell Technologies Capital pairs with Observability 3.0 launch, a new generation of AI-native capabilities including an autonomous AI SRE agent that transforms raw telemetry into operational intelligence, anomaly detection, and LLM observability. The platform already serves over 7,000 organizations with petabyte-scale ingestion; its open-source project has crossed 19,000 GitHub stars. The Series A investment will scale go-to-market motion and customer success operations, with expansion into new cloud regions in US West and EU, plus Microsoft Azure hosting support. The autonomous AI SRE layer eliminates manual triage; LLM observability extends monitoring coverage across frontend, backend, API, network, servers, security and AI workloads in a unified backend using Apache Parquet and DataFusion.

Read the original at openobserve.ai ↗

Vero Security raises $11M seed for AI agent runtime controls

VentureBeat · Nov 18, 2026 · Industry news

Vero Security raised an $11M seed round led by TLV Partners to enforce AI security policies as agents execute. Founded in 2025 by Israeli entrepreneurs Elad Tessler and Daniel Meirov, the startup says several Fortune 500 companies already use its AI Runtime Governor platform.

Why it matters Runtime enforcement via AI Runtime Governor inspects prompts and agent actions at execution time rather than post-hoc review, automating the emerging discipline of AI Security Engineering at enterprise scale.

Vero Security announced $11 million in seed funding led by TLV Partners with participation from GTM Capital, Maple Capital and angels. AI Security Engineering—securing the AI tools and agents employees use—is becoming its own discipline, but most large enterprises assign it to someone already overloaded; Vero automates the ongoing work. Its platform combines continuous discovery, security analysis, and runtime enforcement via an AI Runtime Governor that inspects prompts, tool calls, and agent actions to enforce approved policies during execution, automatically translating security team decisions into controls that update as AI tools evolve. Founded in 2025 by Elad Tessler (CEO) and Daniel Meirov (CTO, both IDF 8200 alumni), several Fortune 500 companies already use the platform. Funding will advance product development, expand coverage across enterprise AI technologies and security workflows, and grow US go-to-market presence.

Read the original at venturebeat.com ↗

Mirendil raises $1B for research agents with 10M token context

Unite.ai · Nov 12, 2026 · Industry news

Mirendil announced a $1 billion Series A led by Kleiner Perkins and Clearlake Capital to make frontier AI research accessible to scientists and engineers via AI agents. The platform supports a 10 million token context window (roughly 500,000 lines of code) enabling agents to retain days of work at once.

Why it matters Integration with Google Cloud's AI Hypercomputer spanning NVIDIA and TPU infrastructure enables research-scale AI workloads with co-designed compute, storage, networking and control planes for larger training loops and agent-assisted model development.

Mirendil, founded by two researchers who left Anthropic in December 2025, raised a $1 billion Series A led by Kleiner Perkins and Clearlake Capital, with co-investors including Battery Ventures, Doerr Capital, Gil & Co, Google, M13, Pear VC and S32. The platform features Fern, a model supporting 10 million token context (500,000 lines of code), and Lor for delegating research work; Mirendil announced a multi-year Google Cloud partnership on August 6, 2026 leveraging TPUs and full-stack NVIDIA AI infrastructure. One researcher and AI agents pretrained Mirendil-Base for under $10 million in compute using custom training infrastructure, model architecture, algorithms and data pipelines built from scratch over PyTorch and JAX. Fern's extended context enables debugging training runs from logs, analyzing complete agent traces and auditing training data at frontier research scale.

Read the original at unite.ai ↗

ExtraHop pitches AI agents for network root-cause analysis

ExtraHop · Nov 17, 2026 · Primary source

AI agents transform IT operations by enabling automated health checks, real-time troubleshooting, and decision-grade intelligence. Agents investigate transaction paths, evaluate dependencies, and present forensic root-cause analysis in seconds rather than hours.

Why it matters Cuts MTTI from hours to seconds using wire-derived records; shifts network operations from reactive ticket-driven to proactive machine-speed orchestration, particularly valuable in multicloud architectures.

ExtraHop details concrete agent capabilities for NOCs: continuous monitoring for traffic degradations, automated health checks across network segments/servers/cloud endpoints, and latency-chain analysis across L2–L7 protocols. The key distinction from earlier AIOps is agents act on evidence-backed diagnostics rather than surfacing alerts. Wire-derived transaction records enable rapid triage where manual troubleshooting consumes hours. For practitioners moving from alert-driven to outcome-driven operations, this shift is structural: agents become force multipliers for high-velocity incident response. Agents can correlate performance events across DNS, HTTP, database, storage, and TLS layers simultaneously—work that previously required coordinating multiple teams and dashboards. The agent can immediately verify whether the network tier is at fault or whether the issue lies in application/database layers, eliminating hours of packet-capture gathering.

Read the original at extrahop.com ↗

Aviz adds conversational agent layer to Spectrum-X AI factory operations

Aviz Networks · Nov 16, 2026 · Primary source

Aviz Networks describes AI NOC as a conversational agentic layer on ONES network orchestration for NVIDIA Spectrum-X AI factories. Agents handle Day 1 tenant and GPU allocation, Day 2 monitoring, audits, and troubleshooting in real time.

Why it matters Demonstrates how agentic operations move beyond static provisioning to handle specialized AI fabric demands: GPU allocation, traffic prioritization, congestion detection, and PFC storm mitigation in high-performance computing require real-time agent-driven decisions.

Aviz Networks' operational model illustrates agentic reasoning for infrastructure supporting AI workloads. Instead of static provisioning followed by manual remediation, the AI NOC uses agents to continuously reason about GPU cluster allocation, tenant isolation, and fabric health in Spectrum-X architectures. The Day 1–Day 2 distinction is important: agents don't just monitor; they actively adjust tenant topology, manage traffic engineering priorities, and detect microbursts in real time. This is production-validated where agentic reasoning directly impacts workload performance—if an agent misallocates GPUs or misses congestion, training jobs stall. For high-performance computing or cloud infrastructure practitioners, this illustrates the boundary between classical orchestration (Terraform, Helm, YANG models) and agentic operations (continuous reasoning under uncertainty). The conversational interface abstracts away multi-layer fabric coordination complexity, allowing operators to reason at business level (allocate 64 H100s for 24-hour training job) while agents handle layer-by-layer validation, isolation policy enforcement, and monitoring. Agents can detect and mitigate congestion patterns that classical queue management misses.

Read the original at aviznetworks.com ↗

Telus abandons two Vancouver AI data centres after local opposition

The Globe and Mail · Nov 15, 2026 · Industry news

Telus pulled out of two planned Vancouver AI data centres expected to open by end of 2026 and 2029, after community protests over environmental, community and housing impacts. The facilities were to house 63,000 Nvidia GPUs and consume up to 163 megawatts of electricity.

Why it matters Withdrawal signals real operational constraints on telco expansion into AI infrastructure amid competing local priorities, impacting Canada's sovereign AI infrastructure ambitions and requiring operators to navigate municipal permitting and community engagement.

One of the two Vancouver locations remained subject to rezoning, while a city councillor motion due for November consideration could impose stricter requirements on data-centre owners, including independent environmental assessments and possible fees. This follows March 2026 announcements by Bell Canada to spend $1.7 billion on a 300-megawatt data centre outside Regina, highlighting divergent regulatory and community reception across Canadian provinces for telco AI infrastructure. The withdrawal underscores tensions between national AI sovereignty objectives and local environmental and zoning constraints—a constraint that will shape how Canadian operators build out the distributed edge compute infrastructure needed for AI workloads. The government remains focused on building domestic computing infrastructure, but execution requires navigating municipal permitting and community engagement.

Read the original at theglobeandmail.com ↗

Most LLM ops agents remain copilots rather than autonomous operators

arXiv (cs.SE) · Nov 18, 2026 · Research

A new survey examines the transition from AIOps pipelines to agentic service operations through three coupled lenses: capability, operational authority, and evaluation evidence. The work audits 22 language-model-enabled systems according to their evaluated behavior rather than intended autonomy, finding most remain copilots or read-only investigators.

Why it matters AIOps and SRE leaders evaluating autonomous remediation claims need this framework to assess actual decision-making authority and evidence quality before deploying LLM-based operations tools in production.

Modern service systems combine rapidly changing dependencies, heterogeneous telemetry, and operational actions whose side effects may amplify incidents. Conventional AIOps research has produced effective detectors, localizers, and diagnostic models, but taxonomies typically organize them by task, telemetry modality, or learning technique—describing what a method predicts without indicating which decisions it may make or what evidence justifies that authority. The audit of 22 language-model-enabled systems reveals that most remain copilots or read-only investigators rather than autonomous agents. This research is directly relevant to SRE teams and AIOps practitioners building or vetting incident management tools; it provides a systematic framework for evaluating whether a tool makes autonomous decisions (and if so, with what evidence) versus purely assisting human operators. The gap between vendor claims of autonomous AIOps and demonstrated authority is significant, and this survey provides practitioners a checklist for assessing actual capabilities before purchasing or integrating such systems.

Read the original at arxiv.org ↗

OpenAI ships GPT-6.2 Astra and Sol with cheaper input

ForkLog · Nov 18, 2026 · Industry news

OpenAI introduced GPT-6.2 Astra and GPT-6.2 Sol on November 17, with Astra for complex professional tasks and Sol for everyday work. GPT-6.2 Astra is 20% cheaper on input and output and 60% cheaper on cached input, using 22% fewer output tokens than GPT-6 Astra.

Why it matters GPT-6.2 Astra replaces GPT-6 Astra as the flagship model with a 1,050,000-token context window, improved document/spreadsheet formatting, and 20–60% cost reductions—directly lowering per-task expenses for production deployment at scale.

On November 17, OpenAI released GPT-6.2 Astra and GPT-6.2 Sol: Astra for complex professional tasks, Sol for everyday work. Pricing drops to $8/$40 per million tokens (from $10/$50), with cached input at $0.40. The new flagship model uses 22% fewer output tokens on OpenAI's internal GDPval v2 benchmark and scores 6.2 points higher than GPT-6 Astra. GPT-6.2 Astra improves document handling—it can build editable tables with revenue forecasts, valuations, and sensitivity analyses directly from requirements. Both models are available immediately in ChatGPT Work and Codex (Plus, Pro, Enterprise, Edu tiers) and via API under identifiers gpt-6.2-astra and gpt-6.2-sol. Pro $200/month users get Ultrafast mode (Sol only); Pro $500/month and eligible Enterprise/Edu accounts get Ultrafast on both models, achieving up to 4.5× standard token generation rate. Knowledge cutoff is May 18, 2026 for Astra, June 2026 for Sol.

Read the original at forklog.com ↗

European Commission proposes accountability rules for autonomous AI agents

Emergent.sh · Nov 16, 2026 · Industry news

The European Commission released a regulatory framework on November 16, 2026, introducing mandatory accountability standards, safety requirements, and transparency obligations for AI systems capable of autonomous decision-making in commercial and public sector applications.

Why it matters Clear regulatory rules may accelerate enterprise adoption by reducing legal uncertainty, though compliance requires investment.

The European Commission formalized its regulatory approach to autonomous AI agents with a framework establishing three foundational pillars: mandatory accountability standards, safety requirements, and transparency obligations for high-risk applications. The proposal requires regular third-party audits for high-risk AI agent deployments and positions the EU as a potential global standard-setter; other jurisdictions are expected to monitor implementation closely. The framework arrives as enterprises grapple with scaling AI beyond isolated pilots—48% of companies had deployed generative AI without formal governance policies as of November 2026, according to Infosys data. The regulation creates both compliance burdens and competitive advantages: organizations that treat governance as foundational rather than bolt-on will navigate conformity-assessment requirements more efficiently. The framework reflects broader EU regulatory momentum with the AI Act's calendar now concrete, forcing enterprises to map AI use cases against high-risk classifications and begin documentation before deadlines arrive.

Read the original at emergent.sh ↗

Nearly half of firms deployed generative AI without governance policy

MIT Technology Review · Nov 16, 2026 · Analysis

Scaling AI beyond isolated pilots requires strong data foundation, clear business outcomes, and enterprise-scale governance. Infosys research found 48% of companies deployed generative AI without formal governance policy, with no documented framework or oversight.

Why it matters Enterprise AI scaling is blocked by governance gaps, not technical capability; absence of policy creates compliance risk backlogs forcing disruptive audits.

Enterprise AI adoption has hit a governance bottleneck. Infosys data shows 48% of companies deployed generative AI without formal governance policies—a critical vulnerability as regulatory frameworks solidify. The core barrier isn't model capability but organizational readiness: enterprises need clear policies defining what agents can do autonomously, how data is used for training, disclosure obligations, and accountability chains. The absence of governance doesn't speed adoption; it creates a risk backlog that forces later, more disruptive intervention. Practitioners building enterprise-wide AI platforms must establish monitoring infrastructure to verify agent behavior alignment with business policy and trustworthiness requirements. SAP and similar platforms are emerging to solve this platform-layer governance problem, making enterprise-wide policy enforcement technically feasible rather than theoretical. Regulatory deadlines (EU AI Act, India DPDP) are now on calendars with conformity-assessment work required before enforcement dates, not after. Organizations operating in multiple jurisdictions face staggered compliance milestones requiring documentation and audit-readiness planning immediately.

Read the original at technologyreview.com ↗
Nothing in today's briefing matches that.

Listening

Podcasts and talks
  • New Relic

    New Relic Autopilot Kubernetes Agent: Domain-Specific AIOps for SREs

    New Relic launched the Autopilot Kubernetes Agent, a purpose-built AI agent for SREs and platform engineers designed to troubleshoot complex clusters and resolve incidents faster without specialized Kubernetes knowledge. It runs more than 90 specialized diagnostic detectors across Kubernetes events, object descriptions, logs, and APM traces, isolating root causes and delivering ready-to-run remediation code.

Vendor Radar

Last 7 days · arrows compare with the 7 before

What changed this week

Last 7 days vs the 7 before

Biggest moves

Trending topics

agent · observability · MCP · Agentic AI · automation · Kubernetes · inference · AIOps · RAG · OpenTelemetry · LLM · SRE

Get the daily briefing

Today's 3 things that matter and every story with why it matters, in your inbox each morning. Free, and you can unsubscribe at any time. Prefer a reader? Follow the RSS feed.