You're reading the Monday, October 19, 2026 edition. Today's briefing →
Toronto
Monday, October 19, 2026No. 127

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·Primary source

    Eleven years of data question whether RPKI destabilizes BGP

    LACNIC/APNIC analysis examines over 11 years of RPKI and BGP data to measure correlation between routing activity and RPKI-related changes, questioning whether growing RPKI deployment could destabilize routing.

    Why it matters Rapid RPKI adoption may create cascading validation failures or misconfigured ROAs; operators must understand if deployment patterns threaten BGP stability at 70%+ coverage.

  2. Routing·Primary source

    Cloudflare cuts HTTP telemetry pipeline from 9GB/s to under 2GB/s

    Cloudflare engineering post details data pipeline optimization: sampling applied upstream, replication rearchitected, aggregations introduced—reducing HTTP dataset from 9+ GB/s to <2 GB/s compressed, achieving ~80% wire reduction.

    Why it matters Sampling moved upstream and aggregation deferred to pipeline ingestion; resilient failure handling routes traffic to surviving regions without backpressure when inter-DC links fail.

  3. Telco·Industry news

    AT&T and Verizon push AI into live RAN processing

    AT&T and Verizon executives described AI-based RAN deployments producing measurable results while emphasizing economic and operational constraints for scale deployment.

    Why it matters Shifts focus from proving AI-RAN works to determining ROI at production scale; defines 'AI worthiness' test where performance gains must outweigh computing cost and complexity.

Today's briefing

What happened, and why it matters

9 stories · 8 topics · Updated 5:01 PM ET

Cisco and NVIDIA ship rack-scale AI fabric to 73,728 GPUs

Cisco Solutions Collateral · Oct 13, 2026 · Primary source

Cisco and NVIDIA released NCP-RA (NVIDIA Collective Communications Platform – Reference Architecture) compliant rack-scale AI factory platform, orderable October 2026, scaling to 73,728 GPUs with Cisco Silicon One N9300 front-end and Spectrum-X N9100 back-end fabrics unified under Nexus One architecture.

Why it matters First vendor-neutral NCP-RA reference architecture from non-NVIDIA partner using proprietary silicon; enables neo-cloud and sovereign cloud operators to deploy standardized, multi-tenant rack-scale AI without NVIDIA-exclusive networking hardware.

Cisco published October 13 technical collateral confirming that its Secure AI Factory with NVIDIA rack-scale platform is orderable from authorized channel partners starting October 2026. This marks the first NCP-RA compliance achievement by an NVIDIA technology partner using its own networking silicon. The architecture combines Cisco Silicon One G300-powered N9300 systems for front-end (scale-up) fabric with NVIDIA Spectrum-X Ethernet–silicon based N9100 series for back-end (scale-out) networking, unified under the Nexus One architecture.

The reference architecture scales to 73,728 GPUs—equivalent to 1,024 Vera Rubin NVL72 racks—deployed as a multi-tenant AI factory. Operational management flows through Cisco Cloud Control, providing unified provisioning, monitoring, and lifecycle management. For sovereign and neo-cloud deployments, this architecture enables independent cloud operators to procure standardized NCP-compliant components without lock-in to NVIDIA-exclusive networking stacks. The platform supports both NX-OS (Cisco's flagship OS) and SONiC (open networking OS), allowing operators to preserve existing automation and skills.

For infrastructure operators: this announces production availability and streamlines procurement. Cisco's position as the first non-NVIDIA vendor with NCP-RA compliance on own silicon signals broader ecosystem competition. Channel partners (Supermicro and others) can now quote Cisco-based AI factories as an alternative to NVIDIA's tightly integrated Spectrum-X platforms, particularly for regions or operators requiring multi-vendor optionality.

Read the original at cisco.com ↗

Eleven years of data question whether RPKI destabilizes BGP

LACNIC Blog · Oct 15, 2026 · Primary source

LACNIC/APNIC analysis examines over 11 years of RPKI and BGP data to measure correlation between routing activity and RPKI-related changes, questioning whether growing RPKI deployment could destabilize routing.

Why it matters Rapid RPKI adoption may create cascading validation failures or misconfigured ROAs; operators must understand if deployment patterns threaten BGP stability at 70%+ coverage.

This analysis addresses a counterintuitive risk: as RPKI adoption accelerates globally, misconfigured ROAs or cascading RPKI validation failures could destabilize inter-domain routing. The research uses RIPE Archive data and RPKI-Flutter—which collects RPKI events from geographically distributed validators at high frequency—to quantify correlation between RPKI state changes and BGP instability. This matters because it challenges assumptions that RPKI is purely additive; operators need to understand whether certain deployment patterns or mass ROA updates create routing churn. With RPKI-valid routing jumping from 54% to 70.3% in one year, driven largely by China's adoption, any systemic RPKI misconfiguration now impacts vastly larger share of global Internet traffic. The practical implication for operators is that RPKI validation enforcement requires careful monitoring of ROA lifecycle and coordination across multiple validators to avoid synchronized failures.

Read the original at blog.lacnic.net ↗

Cloudflare cuts HTTP telemetry pipeline from 9GB/s to under 2GB/s

Cloudflare Blog · Oct 13, 2026 · Primary source

Cloudflare engineering post details data pipeline optimization: sampling applied upstream, replication rearchitected, aggregations introduced—reducing HTTP dataset from 9+ GB/s to <2 GB/s compressed, achieving ~80% wire reduction.

Why it matters Sampling moved upstream and aggregation deferred to pipeline ingestion; resilient failure handling routes traffic to surviving regions without backpressure when inter-DC links fail.

This post details practical strategies for managing exponential data growth in large-scale observability platforms. Cloudflare iterated through multiple phases: broke monolithic schema (2022), adjusted compression (2024), rearchitected replication (spring 2026), decentralized sampling (summer 2026), and introduced aggregations (fall 2026). The key insight is that sampling now happens at ingestion rather than tail-end: Logfwdrs send single copy to centralized receivers, which apply sampling before buffering, carrying only compacted stream. This reduces wide-area transit load orders of magnitude. Aggregation trades granularity for cost by measuring unique IP counts per minute and summing HTTP methods per domain per verb, avoiding pipeline tail-end computation. For SRE teams managing observability at scale, this demonstrates how to handle petabyte-scale analytics by shifting compute leftward; resilience improves because isolated DC clusters continue ingesting without backpressure when backbone links blip, and non-affected regions route around isolated failures automatically.

Read the original at blog.cloudflare.com ↗

Attackers chain three Ivanti Connect Secure flaws for root

Sherlock Forensics · Oct 14, 2026 · Analysis

UTA0576 actor chained three Ivanti Connect Secure flaws (led by CVE-2026-104935, CVSS 8.6) to achieve root on VPN appliances, exfiltrate credentials and AD metadata, and run Kerberoasting attacks. CISA added multiple firewall and VPN flaws to the KEV list.

Why it matters VPN appliance compromise enables lateral movement into directory infrastructure; chained authentication-bypass flaws demonstrate attackers pivoting from edge perimeter to core identity systems via credential harvesting and Kerberoasting.

CVE-2026-104935 is an authentication bypass in Ivanti Connect Secure (CVSS 8.6) allowing unauthenticated remote attackers to run arbitrary commands on the VPN appliance. Once inside, UTA0576 extracted credentials stored on the appliance, exfiltrated Active Directory metadata, ran Kerberoasting, captured LDAP traffic, and used Python LDAP tools to hunt for administrator passwords and BitLocker recovery keys, altering logs to hide the intrusion. Two additional zero-days were chained to escalate privileges. This attack demonstrates how VPN appliances serve as beachheads for lateral movement into core infrastructure. CISA also added CVE-2026-0322 (Palo Alto PAN-OS) and CVE-2026-76519 (Cisco Email Gateway) to the KEV list, indicating multiple firewall and mail gateway flaws are under active exploitation simultaneously, compounding remediation pressure across the perimeter.

Read the original at sherlockforensics.com ↗

Opsin raises $21 million to police enterprise AI agent sprawl

Remio.ai · Oct 15, 2026 · Industry news

Opsin, a San Jose-based AI governance and security startup, raised $21 million led by SYN Ventures and Race Capital to address autonomous software sprawl. The platform monitors AI agents and tools deployed across enterprises, identifying risky components and agent behavior while enforcing policies without blocking legitimate use.

Why it matters Network and SRE teams face new risks from unsanctioned AI agents operating autonomously across infrastructure; Opsin's agent-centric security approach fills a gap traditional security tools cannot address.

Opsin's funding announcement on October 15, 2026 reflects a critical infrastructure gap: enterprises are rapidly deploying AI agents without adequate governance or visibility. Unlike traditional security that monitors network perimeter or endpoint behavior, Opsin targets the new problem: autonomous software accessing systems, data, and integrations broadly. The platform tracks which AI agents are deployed, what skills and components (including model context protocol servers) they rely on, and their actual runtime behavior. For network and infrastructure practitioners, this matters because agents often operate with broad system access and make lateral moves that bypass conventional network segmentation. Opsin claims it can identify potentially dangerous agent activity and enforce policies without disrupting legitimate agent operations. The startup spun up quickly—announced to be addressing a nascent but urgent category in enterprise security. SYN Ventures and Race Capital's backing signals investor conviction that agent governance is as important as the AI infrastructure itself.

Read the original at remio.ai ↗

Roc\u00edo Lorenzo takes over as Telenet chief executive

Broadband TV News · Oct 15, 2026 · Industry news

Rocío Lorenzo becomes CEO of Telenet effective November 2, 2026, succeeding John Porter after 13 years. Separately, Telenet CTO Luk Bruynseels is leaving for a new role at parent company Liberty Global.

Why it matters Leadership transition at major European telecom operator during infrastructure consolidation within Liberty Global's new Benelux group; CTO departure to parent signals organizational restructuring.

Rocío Lorenzo will become CEO of Telenet on November 2, succeeding John Porter after 13 years at the helm of the Belgian telecoms and media operator. Lorenzo joins from Cable & Wireless Panamá, where she served as Executive President and General Manager since January 2024. She will report to Stephen van Rooyen, CEO of Ziggo Group, as Telenet and VodafoneZiggo are brought together within Liberty Global's new Benelux telecommunications group. Porter, 68, will work with Lorenzo for the remainder of 2026 to support the transition. Van Rooyen noted Lorenzo brings more than 20 years of international experience alongside knowledge of the Belgian market. Separately, Telenet group confirmed that chief technology officer Luk Bruynseels is leaving the Belgian operator to take up a new role at parent company Liberty Global.

Read the original at broadbandtvnews.com ↗

Microsoft Agent Framework 1.22 retires AutoGen for Harness agents

Developers Digest · Oct 17, 2026 · Analysis

Microsoft Agent Framework (MAF) at Python 1.22 and .NET 1.25 replaces AutoGen with production-ready Harness agent, functional workflow API, and channel packages for distributed deployment. AutoGen moved to maintenance mode October 2025 with no new features.

Why it matters Framework consolidation signals industry alignment on harness architecture for operations agents. For teams building bespoke infrastructure agents, MAF is now the canonical replacement with 22 Python minor releases in six months post-1.0.

Microsoft Agent Framework is the open-source Python and .NET SDK, with Go in public preview, that replaces AutoGen and the Semantic Kernel agent APIs. The same teams build it. Six months after 1.0 it has shipped 22 Python minor releases and the core .NET packages are at 1.25. The biggest additions are a batteries-included Harness agent, an experimental Python functional workflow API, channel packages for hosting, and resilient background execution. The Harness agent provides a production-grade base with memory, tools, and context management out of the box. The functional workflow API (experimental in Python) introduces stateful, resumable agent chains—essential for long-running operations workflows that must survive network partitions and restarts. Channel packages enable deployment across stdio, HTTP/SSE, WebSocket, and other transports, critical for distributed agent architectures in infrastructure environments. AutoGen was originally developed by Microsoft but placed into maintenance mode in October 2025 with no new features being added.

Read the original at developersdigest.tech ↗

AT&T and Verizon push AI into live RAN processing

RCR Wireless News · Oct 19, 2026 · Industry news

AT&T and Verizon executives described AI-based RAN deployments producing measurable results while emphasizing economic and operational constraints for scale deployment. AT&T uses AI at scale across planning, forecasting, deployment, configuration, and network optimization; Verizon is moving AI into real-time gNodeB processing.

Why it matters Shifts focus from proving AI-RAN works to determining ROI at production scale; defines 'AI worthiness' test where performance gains must outweigh computing cost and complexity.

For AT&T and Verizon, the next stage is less about proving technology works than determining where it creates enough value to justify deployment at scale. AT&T has trialed Ericsson's AI-native link adaptation feature and seen significant improvements in throughput and spectral efficiency. Verizon sees particular potential in moving AI beyond centralized functions into real-time processing at the gNodeB, creating an 'AI worthiness' test where any performance or user-experience improvement must outweigh additional computing cost and complexity introduced into the network. Moving beyond trials requires open architectures, automation, and disciplined lifecycle management across distributed networks. Both operators stress that AI must deliver meaningful ROI to justify scale deployment. This marks an inflection point: the technical proof-of-concept phase is ending and procurement decisions will hinge on disciplined TCO analysis rather than engineering feasibility.

Read the original at rcrwireless.com ↗

OpenTelemetry ships first release candidate process metrics in hostmetrics

OpenTelemetry Blog · Oct 13, 2026 · Primary source

OpenTelemetry Collector Contrib v0.160.0 released October 13, 2026, shipping first Release Candidate process semantic conventions metrics in hostmetrics receiver's process scraper, moving toward stable v1 process metrics.

Why it matters Ops/SRE teams can now standardize on stable process-level observability; RC metrics enable community validation before stable release, accelerating adoption of standardized process telemetry.

OpenTelemetry Collector Contrib v0.160.0 (released October 13, 2026) advances the project's maturity for host-level observability. The System SIG has worked for over a year to stabilize process metrics semantics—the dataset describing CPU, memory, disk, and I/O at the process level. The RC designation signals that the specification is feature-complete and expected to be stable, though community feedback is actively solicited during the RC period. For SREs and AIOps practitioners, this is significant because it means you can now standardize on a single, vendor-neutral semantic for process metrics across your observability stack. The hostmetrics receiver's process scraper now emits metrics using the RC semantic conventions, replacing previous iterations. This aligns with OpenTelemetry's May 2026 CNCF graduation and the stabilization of all three core signals (metrics, logs, traces). Operators deploying OTel can confidently use process metrics in production, knowing that semantic definitions will not break in the transition from RC to v1, reducing future migration overhead.

Read the original at opentelemetry.io ↗
Nothing in today's briefing matches that.

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 · automation · OpenTelemetry · MCP · Agentic AI · Kubernetes · AIOps · RAG

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.