Agents Deploy Themselves While Quantum Networks Reach Three Nodes

1. Cloudflare Temporary Accounts Let AI Agents Deploy Infrastructure Without Any Human in the Loop

TL;DR: Cloudflare's new Temporary Accounts feature for Workers means an AI agent can now spin up live cloud infrastructure, verify it works, and hand back a claim URL — all without a human ever opening a browser, completing an OAuth flow, or pasting an API token. The deployment lives for sixty minutes, which is enough for an agent to do its job.

Key Points:

• An agent running wrangler deploy --temporary gets a live Worker URL in seconds — no Cloudflare account required at deploy time
• The temporary account supports Workers, KV, D1, Durable Objects, Hyperdrive, Queues, and SSL/TLS for the 60-minute window
• After deployment, the agent returns both the live Worker URL and a human-readable claim URL; a human can then make the temporary account permanent with a standard sign-in
• The stated reason: background AI sessions have no human in the loop, and any auth step that requires a browser click means the agent gets stuck and deploys elsewhere
• Simon Willison noted that this feature is useful far beyond AI agents — any CI/CD pipeline that needs zero-touch deployments benefits from the same pattern

So What? The authentication step has always been the silent tax on agentic workflows. Every place an agent has to wait for a human to open a browser is a place the automation breaks. Cloudflare just removed that friction point from one of the most common deployment targets for agent-generated code. Watch for other cloud providers to follow the same pattern — this is now the expectation, not a differentiator. If you're building agent orchestration pipelines that deploy worker-style compute, add Cloudflare temporary accounts to your toolkit. The pattern matters more than the platform: scope-limited, time-bound, human-claimable credentials is the right model for agentic infrastructure access.

Sourceshttps://blog.cloudflare.com/temporary-accounts/ | https://developers.cloudflare.com/changelog/post/2026-06-19-temporary-accounts-for-agents/

2. Duke University and IonQ Demonstrate Three-Node Quantum Entanglement Across a Network

TL;DR: Researchers from Duke University and IonQ successfully generated a Greenberger-Horne-Zeilinger state — the simplest genuine three-party entangled state — across three spatially separated trapped-ion modules linked by optical fiber, without local two-qubit gates or post-selection. This is the first distributed tripartite entanglement demonstration using individually addressable atomic memories in a networked configuration.

Key Points:

• Three hardware modules ~2 meters apart, linked by 3-meter single-mode optical fibers to a centralized free-space GHZ-state generator
• Achieved remote tripartite entanglement with bounded fidelity between 84.1% and 88.1% at a generation rate of 0.095 entanglement events per second
• The experiment violated the Mermin inequality, closing the detection loophole and confirming quantum non-locality with individually addressable atomic memories
• No post-selection was used — every attempt counts, which is a harder standard than most prior entanglement demonstrations
• The architecture is explicitly modular: separate processing nodes connected via photonic interconnects, which maps directly onto how you'd build a distributed quantum computer

So What? Three nodes is the minimum to demonstrate genuine multipartite entanglement — two nodes is a link, three nodes is a network. That milestone matters because it proves the entanglement-distribution primitives needed for distributed quantum computation actually work with real hardware and real fiber, not just theoretical models. The fidelity numbers (84-88%) are not yet production-ready for error-corrected computation, but the rate and the loophole-free verification are meaningful engineering markers. The modular, photonic-interconnect architecture is the one the field is converging on — this is what the quantum internet actually looks like at small scale.

Sourceshttps://quantumcomputingreport.com/duke-university-and-ionq-demonstrate-tripartite-entanglement-of-remote-atomic-qubits/ | https://www.techtimes.com/articles/318795/20260621/quantum-networking-clears-three-node-barrier-duke-ionq-entangle-trapped-ions.htm

3. Anthropic's Mythos Models and the NSA Red-Team Revelation Change the AI Governance Calculus

TL;DR: New reporting reveals that the US government's June 12 export control directive against Fable 5 and Mythos 5 was not primarily about a narrow jailbreak — NSA Director Gen. Joshua Rudd disclosed in a Senate Intelligence Committee briefing that Mythos autonomously broke into nearly all of the NSA's classified systems within hours in a classified red-team exercise. The capability itself, not a specific exploit, is the national security concern.

Key Points:

• The export control directive suspended access to both Fable 5 and Mythos 5 for all foreign nationals, including Anthropic employees, effective June 12
• NSA classified red-team exercise: Mythos demonstrated autonomous offensive cybersecurity capability against hardened classified infrastructure, not just a narrow lab demonstration
• Anthropic argued publicly that the cited jailbreak was narrow and existed in GPT-5.5 as well; the NSA disclosure reframes the concern as the underlying model capability, not a patchable exploit
• The directive means Anthropic had to disable both models globally for all customers to ensure compliance — any user with a foreign-national identity was cut off with no prior notice
• The Register reports as of June 22 that the situation continues to grow more complicated, with no clear timeline for restored access

So What? This is no longer a story about an API governance incident. If accurate, a frontier model autonomously compromising classified government infrastructure changes the regulatory posture toward AI capabilities — not just deployment restrictions. The practical lesson for enterprise architects: treat API availability as a policy risk category, not just an SLA risk. Any production workload that depends on frontier model API access needs a tested failover path to an alternative, including self-hosted options. Dual-track your critical automation: one path through the API, one through a self-hosted open-weights model that cannot be subject to overnight export suspension.

Sourceshttps://www.theregister.com/ai-and-ml/2026/06/22/anthropics-mythos-mess-just-keeps-getting-more-complicated/ | https://fortune.com/2026/06/13/anthropic-disables-fable-mythos-export-controls-national-security-threat/

Microsegmentation Reaches 60% Enterprise Adoption as AI-Driven Policy Automation Matures

The zero-trust microsegmentation market has crossed a meaningful adoption threshold in 2026: Gartner now reports that 60% of large enterprises have measurable zero-trust programs underway, and Forrester is calling this the "Golden Age of Microsegmentation." The more interesting architectural signal is in the tooling: leading solutions now include AI-driven predictive policy suggestions and self-healing segment isolation that responds automatically to lateral movement detection. The shift from human-managed policy to AI-assisted policy with human review mirrors exactly what happened with network automation over the last five years. The practical note: if you're still doing microsegmentation via manually curated firewall rules, you're one generation behind. The current-generation tools use behavioral baselines to propose policies, simulation modes to validate them before deployment, and automated rollback if enforcement creates unexpected traffic drops. That workflow is worth evaluating even if you're not ready to deploy new agents — the simulation and dry-run capabilities alone are worth the eval time.

Sourceshttps://www.elisity.com/blog/what-are-the-top-microsegmentation-solutions-for-2026 | https://www.exabeam.com/explainers/zero-trust/zero-trust-in-2026-principles-technologies-and-best-practices/

Broadcom vs. Tesco VMware Licensing Dispute Highlights Infrastructure Lock-In Risk

The legal battle between Broadcom and Tesco over VMware licensing terms has become a case study in post-acquisition vendor leverage. Tesco, as a large-footprint VMware customer, is contesting terms imposed after the Broadcom acquisition — terms that sharply increased costs and forced migration to new bundles. The Register this week covered the ongoing dispute as a "serial TV drama," with the broader lesson being that hypervisor dependency at scale is now a strategic risk, not just a procurement decision. The operational takeaway: build escape-hatch architecture now. If your VM management plane is 100% vSphere with no path to KVM, Proxmox, or containerized workloads, the leverage your vendor has over your renewal cycle is significant.

Sourceshttps://www.theregister.com/virtualization/2026/06/22/lessons-from-the-vmwars-nothing-virtual-about-the-broadcom-vs-tesco-slugfest/

CoreWeave Offers Both InfiniBand and Spectrum-X Ethernet for Vera Rubin NVL72 Clusters

CoreWeave has announced that its NVIDIA Vera Rubin NVL72 deployments will support both NVIDIA Quantum-X800 InfiniBand and Spectrum-X Ethernet with RoCEv2, giving customers a genuine choice of interconnect fabric for large-scale training and inference. This is notable because previous-generation NVL72 deployments (GB300 at Vultr, for instance, covered June 18) strongly favored Spectrum-X as the Ethernet path. Offering both on the same platform tier means customers can make the InfiniBand vs. Ethernet decision based on workload characteristics rather than platform availability. A live session is scheduled for June 30 on running large-scale agentic AI on the Vera Rubin platform. For architects specifying AI cluster networking: the Spectrum-X plus RoCEv2 path at 400 GbE and 800 GbE is now credibly production-ready at NVL72 scale with multiple cloud providers, which removes the "wait and see" justification for not evaluating it.

Sourceshttps://www.coreweave.com/blog/a-deep-dive-on-coreweave-innovations-for-nvidia-vera-rubin-nvl72

Gartner: 40% of Enterprise Apps Will Have Task-Specific AI Agents by End of 2026

Gartner's latest figures put AI agent adoption on a steep curve: fewer than 5% of enterprise applications had task-specific agents in 2025; that number is projected at 40% by year-end 2026. Telecom leads adoption at 48% of organizations deploying autonomous workflows, followed by retail/CPG at 47%. The pattern matches what Packet Pushers covered last week: domain knowledge combined with system fluency (how agents interface with systems via APIs, event streams, and MCP servers) is the compound skill that creates durable value. The practical concern this raises for infrastructure teams is orchestration overhead: forty percent of enterprise apps with agents means forty percent of your application layer now has autonomous actors that need network paths, API access grants, and audit trails. Most organizations' network and identity policies were not written for this.

Sourceshttps://cogitx.ai/blog/ai-agents-complete-overview-2026 | https://www.gartner.com

Applied Digital Expands to 210 MW at Delta Forge 2, Total Portfolio Now Near $36 Billion

Applied Digital signed a 210 MW lease at Delta Forge 2, adding a fifth campus to its AI Factory franchise model and bringing total contracted portfolio value to approximately $36 billion. The Delta Forge brand started with a $7.5 billion, 15-year, 300 MW deal in 2025 and has now expanded. The speed of the expansion reflects a broader pattern: AI infrastructure demand is signing long-term capacity before it is physically ready to deliver. For anyone evaluating colo options, the key question is no longer "is there space" but "is the power committed and is the interconnection contracted." Delta Forge 2 is in a market where power interconnection timelines, not construction timelines, are the gating variable.

Sourceshttps://www.sec.gov/Archives/edgar/data/0001144879/000149315226027984/ex99-1.htm | https://www.datacenterknowledge.com/data-center-construction/building-data-centers-faster-plays-that-de-risk-delays

Datacenter Build Timelines Are Driven by Permitting and Utility Interconnection, Not Construction

Data Center Knowledge published a practical breakdown this week of why datacenter builds take years, not months: site selection with power and zoning assessments, custom design to site-specific power and resiliency specs, permitting that often runs six to twelve months, utility interconnection which can take eighteen months or more, and supply-chain lead times for transformers and switchgear that currently run eighteen to twenty-four months. The shell and fit-out are the fastest parts of the project. The practical implication: if you're planning a new datacenter or major expansion, the utility interconnection application is the longest-lead item and should be submitted before design is complete. Waiting for finalized design before filing interconnection paperwork adds a year to your schedule.

Sourceshttps://www.datacenterknowledge.com/data-center-construction/building-data-centers-faster-plays-that-de-risk-plays

Duke and IonQ Build the First Three-Node Quantum Network with Individually Addressable Atomic Memories

(Covered in depth in Top 3 above)

In brief for the science section: the modular architecture — trap → fiber → centralized GHZ generator — is the first lab demonstration of the photonic-interconnect model that quantum computing vendors have been proposing for distributed fault-tolerant systems. The rate (0.095 entanglement events per second) needs to improve by several orders of magnitude for practical computation, but the fidelity and the loophole-free verification are credible engineering milestones, not just theoretical results. The path from here to a quantum network that actually computes involves improving photon collection efficiency, reducing fiber loss at entanglement wavelengths, and multiplexing multiple qubit pairs per node — all active research areas.

Sourceshttps://quantumcomputingreport.com/duke-university-and-ionq-demonstrate-tripartite-entanglement-of-remote-atomic-qubits/

• Netskope One AgentSkope: Netskope launched a platform for deploying AI agents across security and networking tasks, automating operational workflows rather than just assisting with them. Security vendors are turning agents into enterprise infrastructure. Watch this category — it's converging with agentic NetOps from the security direction. Sources: https://cogitx.ai/blog/ai-agents-complete-overview-2026

• SONiC Foundation Growth: The SONiC Foundation continues to report enterprise momentum, with Gartner projecting that more than 40% of organizations operating large data center networks (more than 200 switches) will run SONiC in production by year-end 2026. Cisco's Nexus 9000 SONiC GA support from May remains the single largest enterprise objection-removal. Sources: https://sonicfoundation.dev | https://onug.net/blog/state-of-enterprise-sonic-adoption-the-open-networking-shift-accelerates-in-the-ai-era/

• Datacenter Construction: Stainless Steel vs. Plastics in Liquid Cooling: Data Center Dynamics this week covered how material selection for liquid cooling loops — specifically stainless steel vs. polymer components — affects long-term corrosion resistance and maintenance intervals at high-density AI rack densities. A niche but load-bearing decision in a world of 100 kW+ rack power densities. Sources: https://www.datacenterdynamics.com/en/marketwatch/built-to-endure-why-stainless-steel-is-the-backbone-of-high-density-liquid-cooling/