Digital Plumber

OpenTelemetry's May 2026 CNCF graduation represents the project's maturation to the same status as Kubernetes and Prometheus, following independent security audits and formal governance review. The Kubernetes Attributes Processor reached v1.0.0 in September, providing stable, vendor-neutral attribute enrichment for container workloads. Continuous Profiling—based on Linux-perf and eBPF collection—entered public alpha in March 2026 and is now supported natively in the OpenTelemetry Collector and SDKs for Go, Java, Python, and .NET. Over 12,000 contributors from 2,800+ companies have contributed to the project. The Java Agent 3.0 release candidate (v2.32.0) delivered architectural improvements ahead of its October 2026 target release. For network operations teams, OTel's maturity means vendors, observability platforms, and infrastructure tools now have a stable, standardized way to emit telemetry—reducing manual data pipeline work and enabling true multi-vendor, multi-signal correlation in unified NOC platforms.

The Ultra Ethernet Consortium released UE Specification v1.1, advancing Ethernet capabilities for large-scale AI and HPC deployments. Following successful adoption of v1.0 across next-generation AI infrastructure, the v1.1 release focuses on two critical priorities: scale-up networking optimization and enhanced congestion control mechanisms designed specifically for AI workload patterns. The update addresses known pain points in Ethernet-based AI clusters, particularly managing collective operations and reducing tail latency during synchronized training across thousands of GPUs. This follows the UEC's established roadmap around Programmable Congestion Management (PCM) and small-message performance improvements. The v1.1 specification maintains backward compatibility with v1.0 implementations while providing operators running standardized Ethernet fabrics a clearer path to competitive performance parity with purpose-built solutions like NVIDIA Spectrum-X and InfiniBand XDR. Implementation guidance and practical deployment examples are expected to follow in 2026.

Nscale's flagship data center in Glomfjord, Norway, is six months behind schedule, with a floating hotel being towed to house construction workers to accelerate completion. As of July 31, 2026, Nscale operated five active and nine contracted data center sites across 831 MW of active and contracted capacity, but execution challenges are surfacing at flagship deployments. Nscale has a $45 billion Anthropic deal that Anthropic can walk away from if major projects slip, with Microsoft and Anthropic accounting for about 85% of its contracts ahead of the planned IPO. The UK project withdrawal signals deeper grid infrastructure constraints: Microsoft withdrew from what would have been the UK's largest AI data center project, with power expected not until the 2030s, forcing the company to explore alternative energy sources. These real-world construction and power challenges underscore that AI data center networking performance is constrained not just by switch and optics silicon, but by power delivery reliability and construction timeline certainty.

RPKI adoption accelerated sharply in 2026, with routed prefixes protected rising 16.4 percentage points year-over-year from 54.0% to 70.3%. By IPv4 address count, coverage climbed from 53.1% to 65.0%. IPinfo's measurement from September 28, 2026 shows a typical valid prefix reaches 596 vantage points while an invalid one reaches just two, demonstrating RPKI's effectiveness at stopping origin spoofing. China drove most gains, with IPv4 share rising from 42.7% in 2023 to 65.0% in 2026. However, RPKI validates only who may originate a prefix, not the path, and cannot catch attackers staying within published permissions. A Hetzner incident on August 28, 2026 produced an RPKI-valid route for 162.55.80.0/24 inside a permitted /16 to deliver malware. Roughly 16.2% of IPv4 and 20.9% of IPv6 RPKI-covered routes have such loose 'open windows', highlighting the need for complementary mechanisms like ASPA for path validation.

At 21:05 UTC on October 9, a data center provider lost bidirectional network connectivity between its facility and other company data centers, plus outbound internet access. The provider attributed the outage to an error made during modification to DDoS protection configuration. Accounts using that data center could not reach the primary database to look up account details before processing work, halting queue processing entirely. Run state could not be saved because state services lost access to dependencies outside the data center; accounts whose state was stored there could not create, load or save state, preventing progress. Even accounts with no dependency on that data center were affected because their events share an event-processing pipeline with affected accounts. This incident highlights risks of single-provider network changes cascading through shared infrastructure and emphasizes the importance of careful change management for DDoS and network policies that can isolate critical connectivity.

AISLE was founded in 2025 by Ondrej Vlcek, former CEO of Czech antivirus maker Avast, with co-founders from Google DeepMind and Anthropic. The company uses AI to discover previously unknown security flaws, verify findings, and help develop fixes—identified all twelve vulnerabilities in January's OpenSSL release and 18 of 24 in curl's most recent release. This marks a significant move by an LLM company into proactive vulnerability hunting rather than defensive tooling. The deal extends Cohere's European footprint with AISLE's engineering hub in Prague and reflects Cohere's continued investment in Europe. The acquisition aims to boost the combined group's security and threat detection capabilities for enterprise and public sector clients, with private deployment available on customer systems. Financial terms were not disclosed; the deal is a signed letter of intent preceding a definitive agreement.

NinjaOne's announcement reflects the broader ecosystem maturation of MCP as an operational interface for AI. Katana AI uses the MCP server pattern to connect IT operations tools—telemetry, alerts, device history—through one protocol, allowing agents to diagnose and suggest fixes without custom integrations. The product pairs agent-driven analysis with human approval workflows. This is significant for IT operations practitioners because it demonstrates how agentic systems can attach to existing tools (ticketing, ITSM, asset management) without requiring full platform replacement. The MCP server model allows independent IT teams to expose their own tools and data as agent-consumable APIs. As of October 2026, MCP adoption remains pragmatic rather than universal—this announcement signals that enterprise IT vendors are moving from pilots to production deployments of agent-assisted operations, with structured governance and existing-tool reuse as core design principles.

Radware's MCP Server Protection addresses a critical gap in agentic operations: as AI agents gain broader infrastructure access through MCP servers, the MCP protocol itself becomes a new attack surface. Radware positions the solution as observability and protection for the MCP layer itself—tracking which agents connect to which servers, what tools they invoke, and detecting protocol-level abuse. This is directly relevant to network and SRE practitioners deploying agents in production because it bridges the governance-visibility gap that many organizations face when running agentic systems. According to Stacklok's State of MCP survey, nearly half of software companies reported using MCP in production, with 19% reporting broad production use. The product model—cloud-based service or on-premises deployment—allows teams to layer security controls over MCP without replacing agents or infrastructure. For NetOps and SRE teams, this adds necessary auditability when agents execute network changes or incident response workflows.

Context-Mode addresses a practical constraint in agentic operations: LLM context windows fill quickly when agents call multiple tools and receive verbose outputs. By compressing tool results (98% reduction) while preserving critical information, Context-Mode extends the planning horizon for agents operating across many systems. The 17-platform routing via MCP is significant for NetOps and SRE teams because it demonstrates how standardized protocols scale agent workflows across heterogeneous infrastructure stacks. The tool is designed explicitly for autonomous agentic workflows and long-horizon software engineering, with models engineered to run locally. For infrastructure practitioners, this matters because agents that can reason across more steps without refocusing (due to context limits) can handle more complex multi-domain orchestration tasks—for example, coordinating network changes, infrastructure validation, and incident remediation in a single continuous workflow. The approach sidesteps the "agent call depth" limitation that often forces teams to decompose work into serial steps.

OpenSRE addresses a critical gap: while vendors ship agentic SRE platforms with proprietary data models, OpenSRE lets organizations train agents on their own incident data, runbooks, and infrastructure topology. The framework surfaces evidence scattered across logs, metrics, traces, runbooks and Slack threads to resolve production incidents. The significance for SRE and network operations teams is that OpenSRE treats incident response as a learnable task with measurable training data—similar to how SWE-bench does for coding agents. Teams can evaluate agents on their own incident backlog, iterate on prompts and tool definitions, and deploy agents that understand their specific infrastructure. The 60+ tool connectors span monitoring (Prometheus, Grafana, DataDog), incident platforms (PagerDuty, Opsgenie), and infrastructure (Kubernetes, cloud platforms, observability stacks). This is particularly valuable for teams with custom-built or unusual infrastructure where off-the-shelf agents perform poorly. The open-source model also allows integration with NetOps platforms (NetBox, Itential, etc.) via straightforward tool definitions.

At RCR Wireless News's Telco AI Forum 2026, AT&T's Adam Loddeke and Verizon's Cindy Dong outlined how both carriers are moving beyond proof-of-concept into production deployment of AI across the radio access network. AT&T is employing AI across the entire network lifecycle—from capacity planning to field optimization—with a pragmatic stance on whether GPUs are necessary for RAN applications. AT&T has been trialing Ericsson's AI-native link adaptation feature, which embeds machine learning directly into radio firmware to improve spectral efficiency and throughput. Verizon is positioning GPU processing at the network edge for consumer-facing applications like augmented reality rather than inside the radio stack. Both operators emphasize that economic justification—not just technical feasibility—is the gate for wider deployment: additional compute resources, operational overhead, and management complexity must yield measurable performance gains before scale-out. The critical challenge is determining where AI generates sufficient value across thousands of distributed sites to warrant enterprise-wide rollout.

RFC 10055 addresses a persistent gap in ECN adoption: while ECN support exists in most modern operating systems for both TCP and UDP, configuration remains inconsistent and often incomplete. This informational RFC documents how to properly enable ECN on UDP sockets across Linux, Windows, macOS, and other platforms, including API calls, kernel parameters, and testing approaches. For infrastructure teams running low-latency or real-time UDP services (gaming, media streaming, VoIP), ECN enables the network to signal congestion without packet drops, allowing applications to back off gracefully. Network operators benefit because ECN-aware UDP traffic reduces the need for dropping packets during congestion, improving throughput and latency. The guidance is especially relevant for teams deploying QUIC, SCTP, or custom real-time protocols. Practical implementation requires OS kernel support verification, network device support (most modern equipment now signals ECN), and application-layer changes to respond to congestion signals—commonly handled by modern libraries.

DeepSeek released V4.1-Flash-1017 into production on October 17, 2026, paired with a 77.0 Pass@1 score on DeepSWE v1.1. Multiple outlets described the update as delivering broad gains across coding and agentic tasks rather than a narrow improvement on a single test. The model scores 74.2 on DeepSWE v1.1 while charging $0.15 per million input tokens, substantially undercutting closed-source alternatives. The refresh improved performance metrics and cut task runtimes, with one source reporting inference speeds above 237 tokens per second. The release paired the model update with opening paid Search and Fetch APIs at $0.75 per 1,000 calls, expanding the platform's agentic capabilities for practitioners building autonomous systems. For engineers deploying coding assistance and agent loops at scale, the combination of performance (matching or exceeding flagship models on DeepSWE) and price ($0.15 cached input tokens) represents a meaningful shift in the cost-performance frontier for frontier-class capabilities.

Illinois Governor JB Pritzker is setting up a new group within state government to explore more regulations for artificial intelligence as calls grow across the country for new restrictions. The executive order establishes an AI Cabinet made up of experts and representatives from several state agencies that will formulate policy proposals through 2027 and analyze the various risks posed by AI to the safety of Illinois residents and the critical infrastructure residents rely on. This represents a tactical divergence from the Trump administration's push for federal preemption of state AI laws, signaling that state-level governance initiatives continue despite federal pressure. The cabinet's focus on critical infrastructure risk assessment aligns with the White House Super Intelligence task force's mandate to evaluate cybersecurity and national security risks from frontier AI systems. Illinois is following California, Connecticut, Massachusetts, and other jurisdictions establishing formal AI governance structures; the multi-year timeline through 2027 suggests this is a foundational effort rather than emergency response.

In his final executive orders as California governor before his term ends in January, Gavin Newsom issued measures asking for independent verification of artificial intelligence models and established an expert committee to create kill switches for AI agents that could go rogue and cause harm. These orders reverse earlier positions where Newsom had vetoed similar measures, signaling a hardening stance on AI safety as he positions himself for a 2028 presidential run. The independent verification requirement directly parallels the White House Super Intelligence task force's focus on testing and evaluation frameworks, suggesting convergence between state and federal governance priorities. The kill-switch mechanism addresses agent runtime control—a category attracting massive venture funding ($338 million in July 2026 alone across five governance startups). This creates operational urgency for enterprises deploying agentic systems to implement corresponding safety controls and verification protocols, likely becoming table-stakes for enterprise AI procurement.