Oct 6: CISA flags exploited Cisco SD-WAN Manager flaw
Today's 3 things that matter
- CISA flags exploited Cisco SD-WAN Manager flaw
SD-WAN Manager controls enterprise-wide WAN fabric configurations; this hex-encoding flaw enables direct topology manipulation and traffic hijacking across your entire SD-WAN infrastructure. - Ethernet gains ground on InfiniBand across AI fabric tiers
Tomahawk 6 Davisson CPO now competes directly with InfiniBand NDR for tightly-coupled training—Ethernet is no longer just a scale-out choice. - STAMP extension for IP header reflection enters last call
Enables operators to measure how the network actually handles IP headers in flight — reveals whether ECN is marked, path fragmentation behavior, and IP processing variations critical for tuning congestion control and understanding end-to-end path characteristics.
Full stories below, grouped by topic.
Data Center & AI Networking
Ethernet gains ground on InfiniBand across AI fabric tiers
Eremons Notes · Oct 6, 2026 · Analysis
What happened: Broadcom Tomahawk 6 and Ultra chips now target scale-up and scale-out workloads, encroaching on NVIDIA's proprietary fabric turf. Data points to stronger Ethernet adoption across all three scaling dimensions, suggesting lower InfiniBand penetration and shift toward multi-vendor Ethernet ecosystems.
Why it matters: Tomahawk 6 Davisson CPO now competes directly with InfiniBand NDR for tightly-coupled training—Ethernet is no longer just a scale-out choice.
Industry-wide optical component supply constraints are braking growth for Arista and peers near-term, but Broadcom's Tomahawk 6 now targets scale-up workloads previously dominated by NVIDIA proprietary fabric, shifting the baseline toward multi-vendor Ethernet. Optical component tightness enables Cisco capacity recovery as relief valve for medium-tier deployments, though Arista's unified EOS across scale-up/scale-out/scale-across continues reducing operational friction versus whitebox deployments. The analysis reflects H2 2026 reality: Tomahawk 6 Davisson CPO directly challenges the long-held assumption that InfiniBand remains mandatory for frontier training clusters. Stronger Ethernet adoption signals lower InfiniBand penetration going forward, with complexity rising such that integrated platforms (Arista) outperform commodity whitebox approaches in operational efficiency.
Read the original at eremosnotes.substack.com
Network Security
CISA flags exploited Cisco SD-WAN Manager flaw
CISA · Sep 30, 2026 · Primary source
What happened: CISA added CVE-2026-76504 (Cisco Catalyst SD-WAN Manager Hex Encoding Vulnerability) to its KEV Catalog on September 30, 2026, based on evidence of active exploitation. The flaw affects the SD-WAN Manager console with a three-day federal remediation deadline.
Why it matters: SD-WAN Manager controls enterprise-wide WAN fabric configurations; this hex-encoding flaw enables direct topology manipulation and traffic hijacking across your entire SD-WAN infrastructure.
CVE-2026-76504 affecting Cisco Catalyst SD-WAN Manager was actively exploited when added to CISA's KEV catalog on September 30, 2026, with a federal remediation deadline of October 3, 2026. The hex-encoding vulnerability in the SD-WAN Manager—the central console that programs SD-WAN fabric configurations—creates a critical attack surface for network operators. Successful exploitation could allow attackers to manipulate SD-WAN policies, reroute traffic, or inject rogue peers into the fabric, with implications cascading across all connected branch and remote sites. This follows earlier critical Cisco SD-WAN authentication bypass flaws (CVE-2026-20182, CVSS 10.0) disclosed earlier in 2026. Organizations with internet-exposed SD-WAN Manager instances or default credentials require immediate patching or network isolation.
Telco & Cable AI
Bell Canada and Cisco plan sovereign AI infrastructure
SDxCentral · Sep 30, 2026 · Industry news
What happened: Bell Canada and Cisco signed an MOU to collaborate on sovereign AI and digital infrastructure combining Bell's AI Fabric with Cisco's networking, security, and infrastructure technologies deployed within Canada.
Why it matters: Delivers full Canadian sovereignty over infrastructure and operations for critical workloads, addressing CLOUD Act constraints on US-based cloud providers.
Bell and Cisco will combine Bell's AI Fabric with Cisco's networking, security, and infrastructure technologies, including Cisco's Sovereign Critical Infrastructure service. Bell is quadrupling its Saskatchewan data centre facility from 300 megawatts to 1.2 gigawatts, with 300 MW sourced from the provincial grid and up to 900 MW provided by Bell under a Bring Your Own Power model. The $1.7 billion facility began construction with structural frames and foundations already in place. Under the US CLOUD Act, data on American cloud providers—even when physically in Canada—can be compelled by US authorities, making genuine Canadian jurisdiction and control across every technology layer critical for regulated workloads.
Read the original at sdxcentral.com
AT&T commits $3 billion to Corning fibre supply
RCRWireless · Sep 29, 2026 · Industry news
What happened: AT&T and Corning entered a multi-year agreement valued at more than $3 billion with Corning supplying fiber and cable for AT&T's nationwide network expansion as AI drives growing data demand.
Why it matters: Secures 80+ million miles of high-density optical fiber (2027-2032) for dual purposes: residential broadband expansion to 60 million locations by 2030 and long-haul AI data centre connectivity amid strained optical supply chains.
Average AT&T Fiber household now consumes 1 terabyte per month (5x the 2016 level), expected to reach 2-2.5 TB by 2030 as streaming, cloud, and AI applications drive demand. The agreement covers 80+ million miles of high-density optical fiber and connectivity solutions between 2027 and 2032, with infrastructure serving both broadband-to-home expansion and long-haul routes connecting AI data centres. This follows similar major commitments from Verizon, Zayo, Lumen, Meta, and Nvidia around Corning's fiber and optical capacity as AI infrastructure buildout intensifies and tightens global optical supply chains.
Read the original at rcrtech.com
Research, Standards & Industry
IETF publishes Roughtime with timeserver malfeasance detection
RFC Editor / IETF · Oct 5, 2026 · Primary source
What happened: RFC 10049 published an experimental protocol enabling secure rough time synchronization and cryptographic proof of timeserver malfeasance. Clients can detect and report inconsistencies between servers without prior knowledge of current time, using signed responses with nonces and a static configuration of trusted servers and long-term public keys.
Why it matters: Addresses critical bootstrapping problem for time-unaware devices and provides defense against timesource attacks — relevant for infrastructure security where NTP/NTS lack malfeasance reporting.
Roughtime solves two critical problems: enabling devices with no time knowledge to achieve secure synchronization, and detecting when timeservers misbehave. Unlike NTP (RFC 5905) or Network Time Security (RFC 8915), Roughtime requires servers to include a signature over a value derived from the client's nonce-containing request. This binds responses cryptographically to specific clients and enables clients to prove server misbehavior. The protocol uses a "radius" parameter to indicate server certainty about the reported time (e.g., 1,000,000 microseconds = 1-second confidence). Configuration consists of a list of trusted servers and their long-term public keys, which act as roots of trust and ideally remain unchanged throughout the server's lifetime. Operational experience is still needed to evaluate deployment viability, particularly around ecosystem maintenance and malfeasance report processing — the reason for experimental status. Transport is UDP by default with optional TCP support.
Read the original at rfc-editor.org
STAMP extension for IP header reflection enters last call
IETF IESG · Oct 1, 2026 · Primary source
What happened: draft-ietf-ippm-stamp-ext-hdr-15 entered Last Call toward Proposed Standard on October 1, 2026. The draft extends STAMP to allow Session-Reflectors to reflect received STAMP packet IP headers back to senders, enabling measurement of IP-layer path characteristics including ECN marking, IP fragmentation, and header processing.
Why it matters: Enables operators to measure how the network actually handles IP headers in flight — reveals whether ECN is marked, path fragmentation behavior, and IP processing variations critical for tuning congestion control and understanding end-to-end path characteristics.
draft-ietf-ippm-stamp-ext-hdr extends STAMP with a new TLV that allows the Session-Reflector to echo back IP headers from received test packets. This simple mechanism exposes IP-layer behaviors during active measurement. Operators can now observe: whether ECN congestion notification is being marked on the path, if and where IP fragmentation occurs, header processing variations that affect IP options, and per-hop behavior differences. Performance Metrics Directorate early review (completed June 2026) noted this was well-scoped and necessary for comprehensive path characterization. The extension complements RFC 8972 STAMP optional extensions and earlier STAMP augmentations (RFC 9503 for Segment Routing, RFC 8972 for general extensions). Last Call deadline allows final community feedback before advancing toward Proposed Standard status.
Read the original at datatracker.ietf.org
GPU-batched 5G simulator speeds network-in-the-loop AI training
arXiv · Oct 5, 2026 · Research
What happened: Isaac-Net, a GPU-batched 5G New Radio (NR) module, bridges simulation bottlenecks by enabling packet-level 5G network simulation at the speed of physics simulators. The approach batches thousands of simulation environments on GPU, solving the gap between slow packet-level simulators and fast but network-ignorant physics simulators.
Why it matters: Enables massively parallel reinforcement learning in networked robotics by simulating realistic 5G behavior at GPU scale — practitioners running distributed AI workloads with network constraints can now realistically test agent behavior under latency, packet loss, and bandwidth limits.
Robot simulators reduce the network to independent delay per message while packet-level simulators run one scenario per CPU and cannot keep pace with thousands of parallel environments. Isaac-Net, a GPU-batched 5G New Radio (NR) module, advances the uplink of thousands of environments in lockstep with Isaac Lab physics. The paper addresses a critical bottleneck in AI training for networked systems: existing approaches either ignore network realism (GPU physics engines) or cannot scale to parallel training (packet-level simulators). Isaac-Net implements 5G uplink simulation directly on GPU with batched operations, allowing researchers to train agents for thousands of parallel robot environments while maintaining realistic network conditions. The code is open source and available on GitHub. This is directly relevant for practitioners building autonomous systems that must handle real 5G connectivity constraints — from swarm robotics to distributed edge AI inference.
Read the original at arxiv.org
AI Model Providers
Mistral ships 1T-parameter open-weight Large 4 model
LLM Stats · Oct 6, 2026 · Primary source
What happened: Mistral AI released Mistral Large 4, a 1T-parameter open-weight multimodal model trained on 4,000 Grace Blackwell GPUs for general agentic capabilities, positioning itself to compete with American and Chinese frontier models.
Why it matters: Open-weight 1T-parameter alternative enables cost-effective agentic deployments with data sovereignty, reducing reliance on closed APIs for organizations where inference cost and model control matter.
Mistral released Mistral Large 4 (dubbed 'Le Chonk') on October 6, 2026, a 1 trillion-parameter open-weight multimodal model trained on 4,000 Grace Blackwell GPUs. As an open-weight release, practitioners can download, fine-tune, and self-host the model—critical for enterprises requiring data residency compliance or unwilling to route workloads through third-party APIs. The 1T parameter scale positions it at the frontier of openly available models, though practical deployment requires substantial infrastructure. For infrastructure and AIOps teams, this creates a new option for agentic reasoning tasks where cost-per-inference and data residency take priority over access to the latest proprietary frontier models. The multimodal capability extends utility beyond text to vision-based agent observation loops. Key caveats: benchmark performance against GPT-6 and Gemini 4 is not yet published; inference cost and latency characteristics depend on deployment infrastructure chosen; and sustained performance in production agentic workflows remains to be validated at scale.
Read the original at llm-stats.com
AI Industry & Policy
Leaked Anthropic filing shows $518 billion compute commitment
AI Central (Substack) · Oct 5, 2026 · Analysis
What happened: Anthropic's confidential S-1 filing shows 2025 revenue of $4.6B (up 12x from 2024) but operating losses of $8B, with $7.33B in infrastructure spending. The company has committed $518B to compute capacity over a decade and targets a mid-November Nasdaq listing at roughly $2T valuation, implying 31x revenue multiple.
Why it matters: Demonstrates the capital intensity of frontier AI development and divergence between revenue growth and profitability—critical for understanding AI infrastructure financing and IPO precedent-setting.
Anthropic's leaked June IPO documents reveal extraordinary financial realities of frontier AI labs. Quarterly revenue roughly doubled since late 2025, reaching $11.5B in Q2 2026, yet infrastructure spend ($7.33B annually) nearly matches operating losses ($8B). The $518B compute commitment over ten years represents a fixed obligation regardless of utilization, hedging against capacity constraints but locking in staggering capital. Seven co-founders have requested dual-class shares granting 50.1% voting control—mirroring Meta's structure—framed as protection for Anthropic's safety mission against post-IPO pressure. The company targets mid-November Nasdaq listing at $2 trillion, the largest IPO in history, surpassing SpaceX's June 2026 debut at $1.77 trillion. The 31x revenue multiple is floated by bankers and investors, not Anthropic. Dario Amodei has pledged to donate 80% of personal wealth, signaling commitment to stated safety priorities.
Read the original at substack.aicentral.blog
Read this edition on the web · The week in network intelligence · Vendor Radar
Digital Plumber is AI-curated and AI-summarized, with no human review before publishing. Verify before acting on anything here. How it works.