You're reading the Tuesday, October 27, 2026 edition. Today's briefing →
Toronto
Tuesday, October 27, 2026No. 135

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. Security·Primary source

    Arista VeloCloud Orchestrator zero-day exploited for superuser access

    Arista disclosed an actively exploited CVSS 9.9 SQL injection zero-day (CVE-2026-114625) in VeloCloud Orchestrator, the central management plane for SD-WAN networks.

    Why it matters SD-WAN orchestrators are internet-exposed control planes governing thousands of branch and cloud sites. Three zero-days in four months signals sustained targeting; on-premises VCO compromise enables SD-WAN fabric takeover and lateral movement across branch and cloud infrastructure.

  2. DC networking·Primary source

    Meta open sources liquid-cooled 102.4T AI network racks

    Meta shared networking innovations supporting its latest AI models and gigawatt-scale Prometheus cluster at OCP Summit 2026.

    Why it matters Meta's liquid-cooled Tomahawk 6C and Spectrum-6 fabric platforms at 102.4T capacity demonstrate production-ready 800G deployment patterns for hyperscale AI clusters at gigawatt scale.

  3. Research·Primary source

    OpenTelemetry Transformation Language reaches stable version 1

    OpenTelemetry Transformation Language (OTTL) is now v1, becoming stable for traces, metrics, and logs starting with opentelemetry-collector-contrib v0.164.0.

    Why it matters Operators standardizing on OTTL for data pipeline transformations now have a stable, backward-compatible contract for traces, metrics, and logs without concern for breaking changes in 1.x minor releases.

Today's briefing

What happened, and why it matters

15 stories · 10 topics · Updated 11:57 AM ET

HashiCorp adds governed Terraform workflows for AI agents

HashiCorp / IBM · Oct 26, 2026 · Primary source

HashiCorp released new Terraform capabilities enabling humans and AI agents to safely operate infrastructure at scale with governed workflows and connected intelligence through HCP Terraform and Infragraph. Focus on agentic infrastructure operations with task-level automation replaced by trusted governance models.

Why it matters Shifts Terraform from declarative IaC to agent-ready orchestration platform—operators need to understand how governance, state visibility, and MCP integration change their CI/CD change management workflows.

HashiCorp's October 26 announcement signals a strategic pivot: Terraform is evolving beyond static infrastructure provisioning into a foundation for AI-assisted and AI-driven operations. The new capabilities center on three themes: Scale (connected infrastructure intelligence via Infragraph), Safety (governed workflows with built-in approval gates and audit trails), and Agent Integration (MCP-native support for AI agents to propose and execute changes). For network operators, this is significant because it means Terraform state becomes an observable, queryable source of truth that AI systems can reason about—not just execute changes against. The governance layer is critical: as agents become active infrastructure operators, organizations need approval gates, rollback procedures, and audit evidence. HCP Terraform's remote state and policy-as-code now provide these guardrails at the platform level rather than requiring custom middleware. The Infragraph capability generates domain-specific intelligence about infrastructure topology and dependencies, enabling safer what-if analysis. For practitioners deploying network infrastructure with Terraform, this suggests moving toward GitOps-style workflows where proposed changes are validated by policy before agents apply them, with full audit trails for compliance.

Read the original at ibm.com ↗

Datadog ties AI spend to traces and workflow steps

Releasebot · Oct 22, 2026 · Primary source

Datadog Agent Observability connects AI spend to applications, traces, and workflow steps, helping teams pinpoint costly calls and uncover savings opportunities with evidence-based remediation guidance. The feature provides cost-per-token insights and traces expensive LLM calls through application stacks to identify optimization targets.

Why it matters Network and infrastructure teams can now optimize AI workload costs by instrumenting agent spend tracking alongside network telemetry, revealing where application-layer decisions drive network resource consumption.

Datadog Agent Observability connects AI spend to applications, traces, and workflow steps, helping teams pinpoint costly calls and uncover savings opportunities with evidence-based remediation guidance, with cost and token-efficiency insights helping engineers optimize spend while validating changes against quality. Cost Savings Opportunities identify sources of avoidable spend (such as uncached prompt prefixes or conversation history re-sent in full), estimate financial impact, and provide remediation guidance. Agent Observability Experiments enable teams to test baseline and proposed changes against the same dataset, then compare outputs and evaluation results alongside token usage and estimated cost. This bridges AI operations and infrastructure observability, allowing ops teams to correlate AI agent behavior with network impact and optimize accordingly.

Read the original at releasebot.io ↗

Meta open sources liquid-cooled 102.4T AI network racks

Engineering at Meta · Oct 26, 2026 · Primary source

Meta shared networking innovations supporting its latest AI models and gigawatt-scale Prometheus cluster at OCP Summit 2026. IcePack and IceCube are Meta's liquid-cooled network racks featuring 102.4T switch blades with 128×800G ports using Broadcom Tomahawk 6C and NVIDIA Spectrum-6 variants, running FBOSS.

Why it matters Meta's liquid-cooled Tomahawk 6C and Spectrum-6 fabric platforms at 102.4T capacity demonstrate production-ready 800G deployment patterns for hyperscale AI clusters at gigawatt scale.

Meta's OCP presentation detailed production fabric layers for the Prometheus cluster: IcePack and IceCube are the first shipping liquid-cooled fabric racks integrating 102.4 Tbps switch density—each rack houses 128 ports of 800G, achieved via dual ASIC support (Broadcom Tomahawk 6C or NVIDIA Spectrum-6). Running Meta's FBOSS switching stack, these represent the first documented production deployment of Tomahawk 6 silicon at hyperscale. In the scale-up tier, Tahan is Ethernet-based on Broadcom Tomahawk5, engineered for MTIA 300 accelerators at 72 per rack. Meta's ESUN switching software progressed from 1.0 to 1.1, reflecting operational maturity. The strategic significance: both Tomahawk 6 and Spectrum-6 achieve equivalent 102.4T throughput, validating open silicon paths and eliminating vendor lock-in while operating at production density. For practitioners, this confirms both platforms are operationally stable for gigawatt-scale buildouts; liquid cooling is now standard practice for 800G deployment at this density, with real-world FBOSS telemetry available for tuning RoCE congestion control and lossless Ethernet configuration.

Read the original at engineering.fb.com ↗

Power failures drove most Internet outages last quarter

Cloudflare Blog · Oct 27, 2026 · Primary source

Cloudflare Radar examined Q3 2026 global Internet disruptions driven by power failures, natural disasters, and government shutdowns, finding that power dependency of routers, mobile base stations, and data centers was a leading pattern in disruption data.

Why it matters Understanding which failure modes—power, natural disasters, shutdowns—dominate regional outages helps operators prioritize resilience investments and redundancy placement.

The Internet's distributed design depends critically on electricity. When power fails, routers, mobile base stations, data centers, interconnection facilities, and the networks connecting them can fail simultaneously. This post examines Cloudflare Radar's outage detection data across Q3 2026 to identify patterns in what takes the Internet offline. For network operators and AIOps practitioners, this analysis provides concrete data on failure mode frequency: which regional outages stem from power infrastructure collapse, which from natural disasters (flooding, storms), and which result from deliberate government-mandated shutdowns. The findings help practitioners decide whether to invest in backup power, geographic redundancy across regions with different risk profiles, or peering diversity. Cloudflare's perspective is valuable because Radar observes traffic anomalies from millions of domains, not just a single provider's services, giving a broader Internet-scale view than typical vendor postmortems. The report also quantifies recovery time by failure type, which informs incident response planning and SLO design for infrastructure teams managing multi-region deployments.

Read the original at blog.cloudflare.com ↗

CISA confirms active exploitation of two critical PHP flaws

CISA, The Hacker News, BleepingComputer · Oct 27, 2026 · Primary source

CISA confirmed active exploitation of two critical PHP vulnerabilities rated CVSS 9.2 facilitating unauthenticated remote code execution: CVE-2026-109618 (out-of-memory corruption) and CVE-2026-109619 (stale internal request-state). Successful RCE instances were confirmed on honeypot systems October 24, 2026.

Why it matters All versions of PAN-OS and Panorama are affected; any network appliance running PHP web management interfaces (firewalls, load balancers, SD-WAN platforms, orchestrators) faces immediate unauthenticated RCE risk. Federal agencies have until October 29 to patch.

CVE-2026-109618 is an out-of-memory corruption vulnerability allowing unauthenticated remote code execution; CVE-2026-109619 is a stale internal request-state vulnerability with identical impact. Both score 9.2 CVSS. PHP versions 8.5.12, 8.4.27, 8.3.36, and 8.2.35 released October 21, 2026 address these flaws. Adel Karimi at Perplexity confirmed successful RCE on October 24, noting the flaw delivers usable memory disclosure and functional payload execution on single-worker mod_php setups. Palo Alto Networks confirmed all PAN-OS and Panorama versions are affected with no fixed versions available until October 30. Fortinet and Cisco published advisories indicating product investigations are ongoing. PHP 8.1 and older receive no vendor security updates; customers must seek OS vendor backports or migrate to supported versions. Network operators must immediately identify PHP-based management services, restrict access to trusted networks, and prioritize upgrades pending vendor patches.

Read the original at cisa.gov ↗

Arista VeloCloud Orchestrator zero-day exploited for superuser access

Arista Security Advisory 0199, SecurityWeek, Aviatrix Threat Research · Oct 23, 2026 · Primary source

Arista disclosed an actively exploited CVSS 9.9 SQL injection zero-day (CVE-2026-114625) in VeloCloud Orchestrator, the central management plane for SD-WAN networks. Any authenticated user—even read-only accounts—can escalate to superuser operator privileges. This is the third VCO zero-day exploited since July 2026.

Why it matters SD-WAN orchestrators are internet-exposed control planes governing thousands of branch and cloud sites. Three zero-days in four months signals sustained targeting; on-premises VCO compromise enables SD-WAN fabric takeover and lateral movement across branch and cloud infrastructure.

CVE-2026-114625 is an SQL injection vulnerability in VeloCloud Orchestrator On-Prem deployments allowing any authenticated user with basic privileges to escalate to operator-level privileges via low-complexity SQL injection, effectively masquerading as a local superuser. Arista published Security Advisory 0199 on October 22, 2026, noting detection of active exploitation. Fixes shipped the same day across three release trains: VCO 5.2.3.18, 6.1.3.10, and 6.4.2.12. Two attacker-controlled IPs (74.82.197.86 and 38.54.50.151) were observed exploiting the vulnerability before patches became available. Cloud-hosted VCO instances were patched by Arista automatically. The escalation pattern—unauthenticated flaws followed by authenticated privilege escalation—suggests attackers are systematically chaining VCO weaknesses to achieve persistent, high-privilege control over enterprise SD-WAN fabrics. Network operators must immediately upgrade on-premises deployments, audit orchestrator logs for attacker IPs, hunt for webshells and backdoor daemons, and validate all VeloCloud Edge device configurations for unauthorized changes.

Read the original at cisa.gov ↗

Hyperbolic raises $30M to pool idle GPUs globally

Tech Startups · Oct 27, 2026 · Industry news

Hyperbolic raised $30M in Series A-1 funding led by Walden Catalyst (backed by Intel CEO Lip-Bu Tan and former Samsung president Young Sohn) to build a global computing grid connecting idle GPUs across hundreds of cloud providers. Total funding now $50M.

Why it matters Simplifies multi-vendor GPU procurement for AI workloads; operations teams managing distributed AI infrastructure can abstract vendor complexity and access capacity without juggling multiple cloud providers.

Hyperbolic is building infrastructure that aggregates underutilized GPU capacity across hundreds of cloud providers into a single, unified grid, reducing friction for AI companies seeking compute. The Series A-1 round of $30M was led by Walden Catalyst, a venture firm co-founded by Intel CEO Lip-Bu Tan and former Samsung president Young Sohn, with participation from Dell Technologies Capital, General Catalyst, Vista Equity Partners, Cambium Capital, and Salience Capital. The round brings total capital raised to $50M. The problem Hyperbolic addresses is operational complexity: AI teams currently manage relationships with multiple cloud vendors and GPU providers, negotiating capacity and pricing separately. Hyperbolic's grid abstracts that complexity, giving teams a unified API to access compute across providers. This is relevant to infrastructure and NetDevOps practitioners managing AI workload placement, multi-cloud orchestration, and cost optimization—similar challenges that drive interest in cloud abstractions like Kubernetes across network and infrastructure teams.

Read the original at techstartups.com ↗

People Moves

2 stories

NetScout COO Sanjay Munshi resigns effective immediately

SEC Form 8-K · Oct 22, 2026 · Primary source

Sanjay Munshi, NetScout's Chief Operating Officer, resigned from all positions he held with the company and its subsidiaries, effective immediately on October 22, 2026. No reason for departure was disclosed.

Why it matters Unexpected COO departure at major network observability and AIOps vendor; no successor named or transition plan disclosed.

Sanjay Munshi, the Company's Chief Operating Officer, resigned from all positions he held with the Company and its subsidiaries, effective immediately on October 22, 2026. NetScout disclosed the departure via an SEC 8-K filing but provided no rationale or succession plan. Munshi had been promoted to COO only 16 months prior, in June 2025, after serving as Deputy COO. Munshi has more than 30 years of experience in the networking and cybersecurity industry and has led strategic innovations for NETSCOUT's network and application performance, security analytics, and DDoS prevention solutions. The immediate departure and absence of succession clarity marks a significant leadership gap at a company that provides network observability, AIOps, cybersecurity, and DDoS protection to enterprise customers.

Read the original at sec.gov ↗

Potentially AI names co-founder Oliver Yonchev chief executive

AIM Regulatory News Service · Oct 27, 2026 · Primary source

Oliver Yonchev is appointed Chief Executive Officer with effect from 27 October 2026. Arfan Chaudhry was appointed as Chief Product Officer. Sukhveer Sanghera, Co-founder and Chief Architect, stepped down from the Board.

Why it matters CEO transition at AI infrastructure company preparing for workspace product release; aligns technical and commercial leadership structure.

Oliver Yonchev is appointed Chief Executive Officer with effect from 27 October 2026. Oliver co-founded the Company and has served as an Executive Director since its admission to AIM in July 2026, leading the Company's commercial and operational build. Arfan Chaudhry was appointed as Chief Product Officer and Josje Koehof as an Adviser to the Board on AI Governance and Data Strategy. Sukhveer Sanghera, Co-founder and Chief Architect, stepped down from the Board of the Company. These changes, alongside a strategic technology partnership with Juma Labs, are intended to support the delivery of the Company's workspace product, which remains on track for release by the end of 2026. The reshuffle consolidates product and commercial strategy ahead of a critical product launch at the collective AI infrastructure startup, which listed on London's AIM in July 2026.

Read the original at investegate.co.uk ↗

Agentic AI cuts Comcast service contacts by 20%

Deadline / Yahoo Finance · Oct 22, 2026 · Industry news

Comcast is taking a $1 billion Q4 charge including layoffs as part of restructuring its core connectivity business, while agentic AI tools have streamlined customer service operations. Working with AI-agent providers Cresta and Sierra, Comcast has reduced service-contact rates by roughly 20%, eliminating more than 30 million service interactions year-to-date compared with the prior year.

Why it matters Demonstrates measurable ROI from agentic AI in customer service operations; alerts ops teams that AI-driven customer interactions and network automation are tightening staffing and changing ticket workflows.

Comcast flagged a $1 billion Q4 charge for layoffs and restructuring of its core connectivity business during its third-quarter earnings call. Agentic AI tools have streamlined customer service operations, reducing customer service contact rates by roughly 20% and eliminating more than 30 million service interactions year to date versus the prior year, as personal AI agents reshape how consumers discover, purchase and manage services. Comcast is working with AI-agent providers Cresta and Sierra to handle customer interactions. While the company has not explicitly stated that AI will replace all customer-service workers, the combination of workforce cuts and documented 20% reduction in contact volume suggests significant operational shift ahead. This reflects broader industry pattern of telcos deploying agentic AI to handle routine customer interactions, freeing human agents for escalations. For network operations teams, the impact includes changing ticket prioritization, different SLA structures, and potential shifts in how customer-reported issues flow into network operations centers.

Read the original at deadline.com ↗

Singtel virtual agent handles three quarters of calls

Computer Weekly · Oct 26, 2026 · Industry news

Singtel has consolidated years of scattered AI work onto an on-premise sovereign inferencing platform built from scratch, allowing the Singapore telco to scale AI quickly. Its virtual agent Shirley now handles 70-75% of incoming calls without human involvement, with customer satisfaction above 85%.

Why it matters Provides concrete benchmark for agentic AI scaling in telco operations; demonstrates governance model (human-in-loop for network changes) that practitioners should adopt to manage risk while gaining automation ROI.

Gee Bhanawat, chief AI and data analytics officer at Singtel, noted that siloed AI initiatives undertaken by different teams had not achieved enterprise-wide impact, even though the telco had quick wins with AI use cases that delivered value. To support Singtel's goal of becoming "the most intelligent connectivity company in the world", the telco has picked 10 areas where AI can create value, with each use case requiring a clear return on investment for customers, employees, or future business, with the compass being "not AI for the AI activity, but AI for meaningful return." Singtel's AI assistant Shirley answers customer calls and handles 70-75% of incoming calls without any human involvement, with customer satisfaction above 85%. Singtel is rolling out AI agents in stages across its mobile, TV and fibre networks to assist network engineers with root cause analyses and working with Ericsson and Nokia to speed up network recovery, though the AI agents do not operate autonomously due to the risk of making wrong changes to the network. This approach—centralized platform, clear ROI metrics, staged rollout with guardrails—offers a template for other operators planning agentic AI deployments.

Read the original at computerweekly.com ↗

OpenTelemetry Transformation Language reaches stable version 1

OpenTelemetry Blog · Oct 27, 2026 · Primary source

OpenTelemetry Transformation Language (OTTL) is now v1, becoming stable for traces, metrics, and logs starting with opentelemetry-collector-contrib v0.164.0. The v1 release meets the Collector's 'stable' stability criteria for testing, benchmarking, documentation, and telemetry, with a compatibility promise for all of 1.x.

Why it matters Operators standardizing on OTTL for data pipeline transformations now have a stable, backward-compatible contract for traces, metrics, and logs without concern for breaking changes in 1.x minor releases.

As of October 27, 2026, OTTL v1 is available in opentelemetry-collector-contrib v0.164.0 and is stable for traces, metrics, and logs. OTTL syntax and standard functions remain backward compatible, and the Go API for components built on top of OTTL also maintains compatibility. The v1 promotion involved a comprehensive review of every token, type, and feature in the grammar, with all documented language reference elements now frozen for 1.x, covering paths, literals, operators, function calls, and where clauses. New syntax can be added in minor releases, but existing syntax will not be removed or changed in meaning. Some limitations remain: lambda-powered functions (Filter, MapEach, MapKeys, Any, All, Find, Reduce, and When) are still experimental and require the ottl.functions.enableLambda feature gate. For practitioners building operational telemetry pipelines, this signals a mature component suitable for production deployments where consistency across software versions is critical.

Read the original at opentelemetry.io ↗

New RFCs extend PCEP for segment routing multipath

RFC Editor · Oct 26, 2026 · Primary source

RFC 10066 defines PCEP extensions to signal multipath information, associating multiple forwarding paths with a single PCEP LSP for load balancing and redundancy across diverse paths. RFC 10067 extends PCEP support for Segment Routing to group two unidirectional SR LSPs into an associated bidirectional SR LSP.

Why it matters Network operators can now signal multiple equal-cost or weighted paths per LSP and coordinate bidirectional SR paths via PCEP, enabling more granular traffic steering, load balancing, and redundancy at the segment routing layer without requiring separate signaling for each direction.

RFC 10066's primary application is Segment Routing Policy, where a Candidate Path can contain multiple segment lists; current PCEP extensions only allowed signaling of a single segment list per Candidate Path, but these extensions now enable multipath capabilities such as weighted or equal-cost load balancing across paths. RFC 10067 extends the procedure from RFC 10066 to allow grouping two unidirectional SR LSPs into an associated bidirectional SR LSP for both co-routed and non-co-routed paths, utilizing the procedure from RFC 10066 to carry multiple EROs and associated reverse path EROs for an SR LSP. Both RFCs were published October 26, 2026, and represent a maturation of SR control-plane capabilities for operators deploying carrier-grade traffic engineering at scale. These enable finer-grained path optimization and redundancy without requiring parallel signaling channels.

Read the original at rfc-editor.org ↗

AMD to buy World Labs for $8.2 billion

Rio Rundown · Oct 27, 2026 · Industry news

AMD announced an all-stock agreement to acquire World Labs, an AI research lab led by Dr. Fei-Fei Li, for approximately $8.2 billion to integrate spatial-intelligence models into its AI hardware and software development. Dr. Li will join AMD as executive vice president and chief scientist upon closing by end-2026.

Why it matters World Labs develops spatial-intelligence models for 3D environments and robotic learning, positioning AMD to compete with Nvidia in embodied AI and vision-intensive workloads while integrating world-model research at the architectural level.

AMD's $8.2 billion all-stock acquisition of World Labs marks a major consolidation targeting spatial-intelligence capabilities — critical for autonomous systems, robotics, and embodied AI. World Labs, founded by Fei-Fei Li, specializes in 3D environment modeling and robotic learning technologies increasingly essential for industrial automation. Li's appointment as EVP and chief scientist signals AMD's intent to embed this expertise at the architectural level rather than treating it as a bolt-on feature. The deal is expected to close by end-2026 subject to regulatory approval.

For infrastructure practitioners, this matters because spatial AI and vision models are becoming table-stakes for enterprise robotics, autonomous vehicles, and industrial automation. AMD's acquisition directly competes with Nvidia's software ecosystem advantages by combining hardware optimization with foundational model research. The integration positions AMD to offer end-to-end solutions for customers deploying embodied AI, reducing dependency on third-party model providers and improving performance per watt in vision-intensive workloads.

Read the original at riorundown.substack.com ↗

China extends AI rules to autonomous agent behaviour

China Briefing · Oct 26, 2026 · Industry news

China placed comprehensive AI legislation on its 2026 agenda and the amended Cybersecurity Law introduced dedicated AI governance provisions. Regulators issued new measures for anthropomorphic AI services and policy guidance on AI agents, shifting focus from content generation to system behavior and decision-making.

Why it matters China's shift from content-filtering to behavioral governance signals a fundamental regulatory restructuring; enterprises deploying AI agents and autonomous systems in China face new compliance obligations around system behavior, decision-making transparency, and operational risk.

China's 2026 regulatory framework represents a strategic pivot from first-generation AI governance (filtering generated content) toward second-generation oversight focused on system behavior and autonomous decision-making. The amended Cybersecurity Law effective January 1, 2026, introduced dedicated AI provisions, while new rules for anthropomorphic AI companion services (July 2026) and AI agent policy guidance (May 2026) show regulators moving beyond static content controls.

The September 2026 national AI Safety Governance Framework v3 is the most substantive indicator: a dedicated annex on "agent risks" signals China recognizes autonomous AI systems as requiring distinct regulatory treatment from LLMs. For enterprises, this means compliance strategies must now cover system decision-making, agent autonomy thresholds, and operational transparency beyond traditional model cards and fairness audits. State Council legislative work plans explicitly target data governance, computing power oversight, algorithms, and supply-chain security — suggesting forthcoming rules on model training provenance and computational infrastructure allocation. This regulatory velocity likely influences global enterprise AI governance as China becomes a reference jurisdiction for agent-focused oversight.

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

Listening

Podcasts and talks
  • FastNetMon

    NANOG 98 Event Recap: AI Becoming Part of Everyday Network Operations

    According to field observations from NANOG 98 in Miami (October 19–21), AI is becoming part of everyday network operations and infrastructure planning. The conference program reflected topics including AI networking, routing security, RPKI, network automation, quantum-safe networking and global connectivity.

  • Inngest Blog

    Inngest Incident Report: Cross-Datacenter Network and Database Connectivity Failure

    A connectivity disruption between application infrastructure and database impaired database operations; event ingestion services exhausted available memory and restarted, and because startup required a successful database connection, they could not recover while connectivity remained impaired, amplifying the disruption.

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 · OpenTelemetry · automation · Agentic AI · RAG · MCP · Kubernetes · AIOps · NetDevOps · SRE · inference

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.