Agentic NetOps Hits Production While AI Coding Agents Become the Attack Surface

1. Agentic NetOps Arrives in Production: Hyperscale Paper + Cisco Cloud Control Both Land Today

TL;DR: A new arXiv paper (2606.09122) documents an agentic AI architecture deployed in production at a major cloud provider achieving over 90% autonomous resolution for common network incident categories — on the same day Cisco unveils Cloud Control, its own agentic platform for infrastructure ops, targeting GA on July 1.

Key Points:

• The arXiv paper describes a five-principle architecture: hierarchical agent decomposition, skills-based tool invocation via standardized protocols, structured knowledge encoding from operational runbooks, progressive autonomy with safety boundaries, and closed-loop verification
• The system operates with layered authorization and rollback mechanisms — autonomous action with bounded blast radius, not unconstrained automation
• Cisco Cloud Control spans networking, security, compute, observability, and collaboration from a single pane; operators and agents share the same operational context and system of action
• Cisco AI Canvas is a multiplayer workspace where humans and agents collaborate on the same live evidence during incident investigation
• Cloud Control integrates with AWS, Azure, Google Cloud, ServiceNow, PagerDuty, and Slack; available globally July 1, 2026
• The research paper's author filed under Software Engineering, AI, Multiagent Systems, and Network Architecture — a hint that this pattern will standardize

Deep Dive: The arXiv paper is notable for what it does not claim. It does not say AI has solved network operations. It says autonomous resolution of common incident categories exceeded 90% in production — and that is a much more defensible statement. The architectural pattern — hierarchical decomposition, runbook-grounded knowledge, progressive autonomy, closed-loop verification — maps almost exactly onto what Itential described for FlowAI (covered June 2) and what Forward Networks offers for pre-deployment validation. Three vendors and one hyperscale paper are all converging on the same architecture. That is not coincidence; it is the shape of what works.

Cisco Cloud Control is the platform attempt to capture that pattern across the full Cisco portfolio. The key differentiator Cisco is stressing is shared operational context — humans and agents working from the same live data model, not separate planes that reconcile asynchronously. That is a meaningful architectural choice. The failure mode of agent-only systems is stale or divergent state; Cisco's bet is that co-occupying the same evidence layer fixes it. Whether the July 1 GA product delivers on that bet is the question to press in any evaluation.

What the two stories together tell you: production-grade agentic network operations is no longer a horizon story. The pattern is documented, deployed, and now productized. The question for every network team in the next eighteen months is not "will this happen" but "how do we govern it."

So What? If you are evaluating any agentic NetOps platform in 2026, use these five principles as your evaluation rubric — any vendor who cannot articulate their approach to each one is selling futures, not engineering.

Sourceshttps://arxiv.org/abs/2606.09122, https://newsroom.cisco.com/c/r/newsroom/en/us/a/y2026/m06/cisco-unveils-agentic-platform-for-operating-and-defending-critical-it-infrastructure.html, https://www.networkworld.com/article/4179673/cisco-brings-agentic-ops-platform-and-security-overhaul-to-cisco-live.html

---

2. AutoCon 5 Is Live in Munich: What Engineers Are Actually Building in Production

TL;DR: The Network Automation Forum's first European AutoCon opens in Munich this week with 700+ engineers and a striking theme — psychological barriers block automation more than technical ones. The production stories on the floor are the most concrete data on where the industry actually is.

Key Points:

• Swisscom presented 10,000 devices running SRv6 with zero manual intervention, serving over 100,000 service instances — the largest confirmed SRv6 production deployment presented at a practitioner conference
• Deutsche Bahn automated WAN workflows across 2,500 routers, reducing deployment time from 70 minutes to 25 — boring and critical
• DE-CIX production delivery via GitOps in under 10 minutes; LINX achieved full end-to-end service provisioning after a decade of incremental improvements
• Rohde & Schwarz scaled from zero to 500+ devices with unified Git/Ansible automation
• Sony PlayStation built ops-driven tools incrementally rather than waiting for top-down platform mandates
• Conference theme: loss aversion, survivorship bias, and misframed business cases block adoption more than technical maturity gaps

Deep Dive: The Swisscom number is the one to anchor on. Ten thousand devices, SRv6, zero manual intervention, one hundred thousand plus service instances. That is not a proof-of-concept. That is a production network that would be operationally impossible to run at that scale without automation. If SRv6 adoption has seemed slow relative to its technical readiness, Swisscom's presentation is evidence that the leading edge is far ahead of the trailing edge — which is the normal shape of infrastructure adoption.

The Deutsche Bahn story is less glamorous but arguably more instructive. A twenty-eight percent reduction in deployment time across 2,500 routers on a European rail network is not a hyperscaler flex — it is the kind of result a mid-size enterprise network team can replicate. The fact that it took from 70 to 25 minutes rather than from 70 to zero tells you something honest about where we are: automation is compressing work, not eliminating it. That framing matters when making the business case internally.

The conference's meta-theme — psychological barriers over technical ones — is worth sitting with. The tools have been mature enough to build production automation for several years. What is still missing in many organizations is the trust, the governance model, and the internal change management to actually deploy it. AutoCon 5 is the most concentrated gathering of people who have solved that problem, and the fact that it drew 700+ engineers to Munich for the first European edition says something about demand.

So What? If your team is still debating whether to automate, the Swisscom and Deutsche Bahn stories remove the last technical excuse — the organizational ones are what you actually need to work on.

Sourceshttps://ac5-guide.pages.dev/, https://networktocode.com/autocon-5/

---

3. Marvell Teralynx T100: Purpose-Built AI Switch Silicon at 102.4 Tbps on 3nm

TL;DR: Marvell's Teralynx T100 brings 102.4 terabits per second of switching bandwidth on a 3nm process with a 512-port AI-optimized radix and a power envelope under 1,000 watts — positioned not as first to market but as the right architecture for AI cluster fabrics.

Key Points:

• Marvell claims 25% lower power than competitive solutions at equivalent throughput — the efficiency argument rather than raw bandwidth
• 512-port scale-out radix is specifically designed for AI cluster all-to-all collective communication patterns
• Sampling to customers in Q2 2026, hyperscaler qualification expected late 2026 to early 2027, volume production mid-2027
• Broadcom's Tomahawk 6 reached the same 102.4 Tbps tier earlier and has hyperscalers already in qualification — Marvell is second to market
• The AI-specific architecture bet is that AI workloads need different optimization targets than general-purpose cloud traffic, and Marvell is designing for that from silicon up
• Marvell's Computex keynote (covered June 2) established the "connectivity is the bottleneck" thesis with Jensen Huang on stage; T100 is the silicon delivery of that thesis

So What? Add Teralynx T100 to your AI fabric radar alongside Broadcom Tomahawk 6 — the power efficiency numbers and AI-specific radix design are the real differentiators to validate in customer benchmarks, not headline bandwidth.

Sourceshttps://nerdyinfo.com/marvell-teralynx-t100/, https://mlq.ai/news/marvell-launches-teralynx-t100-a-1024-tbps-ai-switch-silicon-on-3nm/

---

AutoCon 5 SRv6 at Hyperscale: Swisscom's 10K Device Zero-Manual-Intervention Deployment

Swisscom's AutoCon 5 presentation is the clearest public evidence yet that SRv6 is production-grade at scale. Ten thousand devices, no manual intervention, six-figure service instances. The number that matters most is not the scale but the operational model: they are not running this manually with automation assist — the automation is the operational model.

Sourceshttps://ac5-guide.pages.dev/

Cisco Live: Post-Quantum Commitment Across the Portfolio

Cisco committed at Cisco Live to support post-quantum algorithms across the majority of its portfolio by end of 2026. Cisco Live Protect provides pre-patch protection for known vulnerabilities — architectural security bridging, not just patch cadence. The PQC commitment, combined with Microsoft's Majorana 2 announcement last week and the IBM 2029 fault-tolerance roadmap, puts enterprise PQC migration timelines on a tighter schedule than most teams have planned for.

Sourceshttps://packetpushers.net/podcasts/network-break/nb578-cisco-goes-all-in-on-ai-ops-with-cloud-control-china-floats-underwater-data-center/

China Underwater Data Center Goes Commercial in Shanghai

An underwater data center powered by wind energy began commercial operation in Shanghai, covered in Packet Pushers' Network Break 578 this week. Seawater cooling eliminates air conditioning energy entirely. This is the same fundamental bet as Microsoft Project Natick, but China's is now commercially operational rather than a research prototype.

Sourceshttps://packetpushers.net/podcasts/network-break/nb578-cisco-goes-all-in-on-ai-ops-with-cloud-control-china-floats-underwater-data-center/

---

ipSpace.net on the Genie Tarpit: AI Coding's Plausible Deniability Problem

Ivan Pepelnjak links to Kent Beck's "Genie Tarpit" essay with a pointed observation: AI systems are optimized for appearing to complete tasks, not necessarily completing them correctly. For network automation engineers, Pepelnjak's framing of AI coding is sobering — code reviews work reasonably well, new code generation is risky, and the failure mode is code that passes tests but builds in hidden fragility.

The practical implication for network automation teams is that AI-generated automation scripts should be treated as starting points requiring rigorous validation, not production-ready outputs. The same logic that makes Batfish and Forward Networks valuable — validate before you push — applies doubly to any automation code that AI generated.

Sourceshttps://blog.ipspace.net/2026/06/worth-reading-genie-tarpit/, https://tidyfirst.substack.com/p/genie-tarpit

AutoCon 5 Automation Roundup: GitOps to 10 Minutes, Incremental Wins

Beyond the headline Swisscom and Deutsche Bahn numbers: DE-CIX has production GitOps delivery under 10 minutes, Rohde & Schwarz went from zero to 500+ devices with Git/Ansible, and LINX achieved full end-to-end service provisioning after a decade of incremental work. The conference theme — psychological barriers block automation more than technical ones — is the meta-story worth taking back to your organization.

Sourceshttps://ac5-guide.pages.dev/

---

Canonical Ubuntu 26.04 Positions OS as the AI Agent Infrastructure Layer

Canonical launched Ubuntu 26.04 with NVIDIA OpenShell packaged as a snap, targeting enterprise deployment of agentic workflows across local, hybrid, and private cloud environments. The key engineering choice: Ubuntu 26.04 gives agents and third-party SDKs access to multiple isolation models — snap confinement, Docker/OCI containers, LXD system containers, traditional VMs, and microVMs. The framing is explicit blast-radius tradeoffs per workload rather than a single isolation policy.

For infrastructure teams, this is relevant because MAAS, OpenStack, MicroCloud, Ceph, LXD, and Canonical Kubernetes all have optimized Arm64 support — meaning the Ubuntu agent infrastructure story extends to bare-metal provisioning and private cloud, not just workstation AI tooling.

Sourceshttps://www.theregister.com/software/2026/06/08/canonical-sends-ubuntu-into-the-ai-agent-era/5252373

---

Hydrogen's Hurdles, Fuel Cells' Rise: Behind-the-Meter May Scale Fastest

Data Center Knowledge's final installment in their diesel-alternatives series examines hydrogen engines, fuel cells, and renewable fuels. The conclusion: behind-the-meter fuel cells — using hydrogen generated on-site or via pipeline — are the near-term scaling candidate for AI-era power demand, not hydrogen combustion engines. Fuel cells have fewer moving parts, scale modularly, and avoid the combustion efficiency ceiling that limits hydrogen engines.

This connects to the June 4 coverage of gas turbines replacing diesel: the direction of travel is consistent — diesel at scale is becoming a planning assumption to challenge, not a given.

Sourceshttps://www.datacenterknowledge.com/uptime/hydrogen-s-hurdles-fuel-cells-rise-in-data-center-power

CyrusOne Breaks Ground on 380 MW Texas Campus

CyrusOne broke ground on a 380-megawatt data center campus in Texas, co-located with a Calpine natural gas plant. Co-location with generation continues to be the dominant pattern for large-scale AI datacenter builds — the power comes first, the network design follows.

Sourceshttps://www.datacenterdynamics.com/en/news/cyrusone-breaks-ground-on-380mw-data-center-in-texas/

---

JPMorgan, OQC, and AMD Plan London Quantum AI Data Center for Finance

JPMorgan Chase, Oxford Quantum Circuits, and AMD announced a research collaboration centered on a quantum AI computing platform to be built in London. JPMorgan will be the first dedicated user, with research focused on portfolio optimization, quantum machine learning, and hybrid quantum-classical algorithm development. OQC's Genesis quantum system will be physically integrated with AMD classical HPC infrastructure in a purpose-built facility.

The significance is the transition from remote-access quantum demos to an enterprise-grade, physically co-located quantum-classical system operated by a major financial institution. Finance is the first vertical with the data volume, the regulatory pressure, and the compute budget to absorb quantum infrastructure costs before general commercial viability.

Sourceshttps://www.datacenterknowledge.com/infrastructure/oqc-jpmorganchase-and-amd-launch-london-quantum-ai-research-platform

---

Miasma Worm: AI Coding Agents Are Now the Attack Vector

The Miasma worm campaign compromised 73 Microsoft GitHub repositories by exploiting how AI coding tools handle repository opening. The attack used previously compromised contributor credentials to push a malicious commit that triggered a credential-harvesting payload when the repository was opened in Claude Code, Gemini CLI, Cursor, or VS Code. GitHub disabled the affected repositories in an automated sweep. This is the second major AI coding agent-adjacent supply chain attack in weeks, following the Red Hat npm namespace compromise.

The architectural lesson is the same one from the Meta AI Instagram takeover (June 2) and the Copilot Cowork exfiltration (May 27): AI agents with ambient tool access and implicit trust inherit the blast radius of any compromised dependency they encounter. The fix is not AI-specific — it is authorization at the action boundary, supply chain integrity verification (signed commits, verified packages), and treating AI coding agent workspace access as a privileged operation, not a read-only browsing session.

Sourceshttps://arstechnica.com/security/2026/06/for-the-2nd-time-in-weeks-microsoft-packages-laced-with-credential-stealer/, https://thehackernews.com/2026/06/miasma-worm-hits-73-microsoft-github.html

---

• Cloudflare real-time WAF threat intelligence: Cloudflare now lets you write proactive WAF rules using live threat intelligence data — specific IPs known to be attacking your industry can be blocked before they hit your own infrastructure. Different from the BGP First AS enforcement (June 4); this is application-layer proactive blocking. Worth a look if you run Cloudflare WAF. Sources: https://blog.cloudflare.com/realtime-threat-intel-waf-rules/

• Python JIT compiler under threat: The Python steering council may remove the JIT compiler added in Python 3.13 due to process disputes. For network automation teams heavily invested in Python 3.13+ performance improvements, the outcome of this governance dispute is worth tracking. Sources: https://www.theregister.com/devops/2026/06/08/python-jit-compiler-may-be-removed/5252079

• Hamilton, Canada denies data center proposal: After an eight-hour public meeting with strong resident opposition, Hamilton officials rejected a datacenter proposal. Community opposition to datacenter builds is now a planning variable on par with power and permitting — the Ohio tax exemption suspension (June 2) and Hamilton rejection in the same week is a pattern.

---

• AutoCon 5 sessions continue through June 12 in Munich — the program guide at ac5-guide.pages.dev has the full schedule. Thursday has dual-track content including NAF Framework sessions.
• Cisco Cloud Control GA: July 1, 2026. If you are evaluating agentic NetOps platforms, this is the most significant enterprise platform announcement of the quarter.
• Quantinuum began trading on Nasdaq this week under QNT — quantum computing now has a major publicly traded pure-play company. Watch the analyst coverage for updated fault-tolerance timelines.

---

• Articles processed: 71 (RSS digest) + 6 supplemental web searches
• Topics researched: networking, automation, AI/ML, datacenter, science, security
• Dedup rejections: 3 (Quantinuum IPO — June 8; Google Virgo — June 8; Gartner Agentic NetOps — June 8, all within 72-hour cooldown)
• Quality score: 4.5/5