Plumbing the
information age

FastNetMon's August 2026 community newsletter addresses core challenges in network operations at scale. The month's focus includes product updates to FastNetMon's core detection engine, the launch of Netomics (a new analytics companion), and hands-on work around 100Gbps DDoS simulation—a practical concern for any operator managing modern data center or ISP-scale infrastructure. The newsletter also reviews botnet research trends and historical context on network infrastructure evolution. For practitioners, this matters because detection and remediation at multi-terabit scales requires rethinking both infrastructure design (can your out-of-band network handle simultaneous updates across thousands of devices?) and operational workflows (how do you validate a mitigation didn't introduce asymmetric routing?). The coverage is practitioner-focused, not vendor marketing—addressing real operational constraints operators face when automation runs at carrier and hyperscaler scale.

Cisco announced on August 25 (reported August 28) an expansion of its Secure AI Factory with NVIDIA to deliver industry-leading rack-scale, liquid-cooled AI infrastructure. The solution combines Cisco networking with NVIDIA AI infrastructure and Supermicro compute systems, offering both liquid-cooled and air-cooled options for dense GPU deployments. Cisco will be the first NVIDIA technology partner to deliver an NVIDIA Cloud Partner (NCP)-compliant reference architecture built on partner-developed networking systems, spanning both Cisco Silicon One and NVIDIA Spectrum-X Ethernet switch silicon, unified by the Cisco Nexus One architecture. The announcement includes a choice between Cisco NX-OS (industry's most deployed data center OS) or SONiC, plus Cisco Validated Infrastructure Services (CVIS)—end-to-end solution design validation aligned with NVIDIA's NVIS methodology. This removes thermal and power constraints that previously kept trillion-parameter training and high-throughput inference out of reach for enterprises, with unified management delivered through Cisco Cloud Control and Cisco Nexus One for network management. The partnership directly addresses infrastructure supply chain constraints and operational complexity that enterprises face when deploying large GPU clusters, potentially shortening time-to-value realization by eliminating months of manual configuration typical in multi-vendor AI cluster deployments.

The Model Hardware Standard addresses a critical friction point in AI-assisted scientific research: integrating autonomous agents with diverse hardware systems. Anthropic, in collaboration with HHMI Janelia Research Campus, built MHS as an open specification that works like a universal driver layer for physical devices. Rather than requiring custom software for each instrument, MHS enables standardized commands and auto-generated reference files detailing operational parameters and safety limits. Early test results are substantive: Genentech automated a protein assay across a liquid handler, robotic arm, and plate reader; QuEra's quantum computer laser stabilization improved from 58% to 99.3% success rate; Tetsuwan environmental researchers used MHS to analyze creek water pollution. The standard is initially available through research preview to select organizations spanning science, robotics, and manufacturing, with Anthropic planning open-source release post-preview. Hardware partners including Universal Robots, Tecan, Danaher, Doosan Robotics, and AWS are integrating MHS support. Critical caveat: current models still require human oversight on physical failure modes; this is not fully autonomous lab operation but rather dramatically faster integration of human-in-the-loop workflows.

The pilot involved the Singapore AI Safety Institute, OpenMined, AVERI, and MLCommons testing Gemini 2.5 Flash Lite against confidential prompts inside a cryptographically secured environment. AVERI used reserved prompts from MLCommons' AILuminate safety benchmark covering cyberattacks, chemical hazards, hate speech, and self-harm, while Singapore's AISI deployed context-specific harmful content tests. The structural problem being solved is fundamental: conventional model evaluation required either exposing test questions to the model provider (compromising benchmark integrity) or exposing proprietary model weights to evaluators. Confidential computing enclaves eliminate this binary. The approach has real implications for regulatory testing: national safety institutes can now run pre-deployment assessments under arrangements that are cryptographically verifiable rather than confidentially negotiated, lowering barriers for jurisdictions unable to share sensitive evaluation material with US companies. However, the pilot involves only a lightweight model and four partners, so widespread adoption depends on whether frontier labs submit their largest models to identical conditions.

Singtel's AI transformation demonstrates practitioner-grade progress beyond pilot stage. In Singapore operations, the telco has deployed AI-enabled personalized customer engagement systems that reduced churn and improved satisfaction ratings. Network operations show measurable gains: faster recovery from faults and efficient deployment of new applications through IT and network automation. Optus (Singtel's Australian subsidiary) has made AI available to all employees, improving workforce productivity and achieving faster network fault identification. The operator's Nxera Digital InfraCo datacentre business now operates 62 MW and is scaling to 230 MW (120 MW Singapore, 110 MW across Malaysia, Indonesia, Thailand) by end of March 2027. Singtel's Chief AI Officer Geet Bhanawat stated the company expects hundreds or thousands of agents working across customer and network operations within a year. The operator is building a sovereign AI platform using RE:AI cloud service to give enterprises local control over GPU capacity—critical for regulated workloads. Singtel has set an internal fault detection target of 1-minute detection, 5-minute root cause identification, and 60-minute restoration, with plans to further reduce this. This contrasts with most telcos still in trial phase; Singtel is operationalizing at scale across a multi-stack vendor environment (Ericsson, Nokia, Huawei, ZTE).

AccuKnox announced AgentZ on August 27, 2026, bringing agents, execution environments, tools, workflows, permissions, and governance into a single platform for enterprise deployment. The platform follows a model of Organizations, Workspaces, Agents, Workflows, and Sandboxes, with workflows using agents for computation, sandboxes for isolation, skills for reusable capabilities, credentials injected at runtime, and triggers determining when workflows run.

For founders and operators building internal agents, AgentZ provides enterprise controls and deployment options (SaaS, on-prem, or air-gapped) that security teams expect, reducing time-to-production and the governance gap that often blocks agent rollouts. This is substantive for enterprises moving AI agents from experimental pilots into production workflows, particularly in regulated industries where audit trails and access controls are non-negotiable. The platform applies Zero Trust and least privilege through kernel-enforced runtime controls, focusing on controlling what agents can actually do, not just detecting what they do.