Amaze Networks — Morning Briefing

---

1. Claude Mythos Autonomously Chains Four Zero-Days — Anthropic Deploys It Defensively First

TL;DR: Anthropic's most capable model, Claude Mythos Preview, autonomously identified and exploited a 17-year-old remote code execution vulnerability in FreeBSD with zero human involvement after the initial prompt. It found 271 vulnerabilities in Firefox alone. Instead of releasing it broadly, Anthropic launched Project Glasswing — giving access to 50 partners including Apple, Google, Microsoft, AWS, CrowdStrike, and NVIDIA, backed by $100 million in usage credits for defensive use.

Key Points:

• Mythos Preview achieved 181 working Firefox exploits out of several hundred attempts, plus register control on 29 more — compared to Opus 4.6's 2 out of several hundred
• The model autonomously chained four vulnerabilities into a single browser exploit, obtained local privilege escalation by exploiting race conditions, and wrote a full remote code execution exploit for FreeBSD's NFS server granting root access
• These capabilities were not explicitly trained — they emerged from general improvements in code, reasoning, and autonomy
• Mozilla shipped Firefox 150 patching 271 Mythos-discovered vulnerabilities
• Project Glasswing gives 12 named launch partners (Amazon, Apple, Google, Microsoft, CrowdStrike, NVIDIA, Palo Alto Networks) plus ~40 additional critical infrastructure maintainers early access
• Anthropic is committing $100 million in usage credits and $4 million to open-source security organizations

The Deep Dive: The significance here is less about the model and more about the architectural moment. Autonomous vulnerability discovery has been a theoretical concern for years — the argument being that AI capable enough to find and exploit zero-days would arrive faster than defenders could adapt. Mythos Preview appears to be that model. The fact that Anthropic is deploying it in a controlled, defender-first posture is the right call architecturally, but it also signals that the asymmetry between offensive and defensive capabilities is closing fast.

What makes Mythos different from previous "AI found a bug" stories is the autonomy and chaining. Finding a single bug is table stakes. Autonomously chaining four vulnerabilities into a working exploit — without a human reviewing each step — means the model is doing reasoning about the attack surface, not pattern-matching on known CVE types. That's a different capability class.

For network and infrastructure engineers, the immediate implication is direct: if you maintain any open-source software or critical infrastructure code, Project Glasswing access is worth pursuing now, before Mythos-class capabilities become more broadly available. The second implication is architectural: if AI agents can autonomously discover and chain vulnerabilities in code, your attack surface just grew to include every piece of software your agents interact with or generate.

So What? Do This: If you have open-source or critical infrastructure code, apply for Project Glasswing access at anthropic.com/glasswing before this capability proliferates. For infrastructure teams: re-examine what software your AI agents can touch — if agents read or execute code from external sources, that surface is now a higher-priority target. The broader lesson: the "defender advantage" from AI tools is real, but only if you use them before the attackers do.

SourcesAnthropic Project Glasswing | red.anthropic.com Mythos Preview | Foreign Policy Analysis | Engadget | The Hacker News

---

2. QuEra/Harvard/MIT Achieve Teraquop Quantum Error Correction with 2:1 Physical-to-Logical Qubit Ratio

TL;DR: A collaboration between QuEra Computing, Harvard, and MIT has demonstrated quantum error correction at a 2:1 physical-to-logical qubit ratio — encoding 1,156 logical qubits in 2,304 physical qubits — entering the "Teraquop" regime of one error per trillion logical operations. This compresses the hardware requirements for fault-tolerant quantum computing by two to three orders of magnitude.

Key Points:

• Achieved using quantum Low-Density Parity-Check (qLDPC) codes co-designed for neutral-atom hardware, with non-commuting affine permutation matrices for syndrome extraction
• The [[2304, 1156, ≤14]] code has an encoding rate of 0.502 — more than half the physical qubits are doing useful logical work
• Per-logical-per-round error rate: approximately 1.3 × 10^-13 (Teraquop threshold = 1 error per trillion operations)
• Hardware advantage: neutral-atom arrays move qubits in parallel via Acousto-Optic Deflectors, enabling constant-time syndrome extraction that aligns with the qLDPC code structure
• Teraquop threshold is the estimated requirement for cryptanalysis-relevant algorithms like Shor's for RSA
• Previous estimates required ~1,000 physical qubits per logical qubit; this result suggests a path at ~2 physical qubits per logical qubit

The Deep Dive: This result lands differently from the steady stream of "new qubit record" press releases. The 2:1 ratio is not just a hardware milestone — it's a resource scaling argument. If you previously estimated needing 10,000 logical qubits for Shor's algorithm at Teraquop fidelity, and assumed 1,000 physical qubits per logical qubit, you were sizing a 10 million physical qubit machine. At 2:1, the same computation runs on roughly 20,000 physical qubits. QuEra has already demonstrated ~1,000-qubit arrays; the Caltech team (covered Tuesday) showed 6,000. The gap between "experimental demonstration" and "cryptographically relevant computation" just got a lot shorter.

The qLDPC approach also matters architecturally: it enables higher logical qubit density, which means smaller physical machines for the same computational power. Combined with Monday's Caltech/ETH Zurich result showing 5 physical qubits per logical qubit (a different code family), there's now a convergence of evidence that the hardware requirements for fault-tolerant QC are substantially lower than the field assumed even 18 months ago.

So What? Do This: If you've been tracking post-quantum cryptography migration as a "2030 or later" problem, this is a signal to accelerate your timeline. The two questions that matter: where is RSA or ECDH in your infrastructure, and what's your migration path to NIST PQC standards (ML-KEM, ML-DSA)? Start the inventory now. It doesn't take a fault-tolerant quantum computer to motivate harvest-now-decrypt-later attacks — those are already happening.

SourcesQuantum Computing Report | The Quantum Insider | QuEra Blog

---

3. Cloudflare Declares "Moving Past Bots vs. Humans" — Intent and Behavior Replace Identity

TL;DR: Cloudflare published a significant architectural position paper arguing that the bot-vs-human binary is obsolete. With AI agents, zero-trust proxies, and accessibility tools accounting for growing fractions of "human" traffic, Cloudflare is shifting to intent and behavior modeling — and introducing an "Agent Readiness" score for sites to understand how well they support legitimate AI agent traffic.

Key Points:

• Cloudflare observes "wanted bots" and "unwanted humans" — the identity axis is no longer the right classification dimension
• The new model focuses on intent and behavior signals: what is the client trying to accomplish? Does it match the site owner's interests?
• Agent Readiness score: a per-site metric indicating how well the site supports legitimate AI agent interactions (structured data, API availability, crawl permissions)
• 32% of Cloudflare network traffic is now automated (covered previously) — this paper provides the architectural rationale for how to handle it
• Zero-trust proxies used by enterprises for employee traffic are specifically called out as a pattern that breaks human-detection heuristics

So What? Do This: Review your bot management rules for false-positive rates against enterprise zero-trust proxies and AI agent traffic. If you're blocking legitimate automation (your own or partners'), the cost is real. Check your Agent Readiness score and consider whether exposing structured API access would reduce scraping pressure while enabling legitimate integrations.

SourcesCloudflare Blog — Moving Past Bots vs. Humans

---

Cloudflare Intent-Based Traffic Architecture

Covered in Top 3 above

AES-128 Post-Quantum Status: Not What You Think

Cryptography engineer Filippo Valsorda published a detailed rebuttal to the persistent claim that AES-128 is "broken" by quantum computing. The short version: Grover's algorithm halves the effective key space, but 64-bit quantum security is still computationally infeasible at any realistic quantum scale — breaking AES-128 would require a quantum computer running for thousands of years even with perfect error correction.

Why it matters architecturally: A significant portion of enterprise PQC migration planning is incorrectly prioritizing symmetric cipher replacement. The real risk is asymmetric cryptography (RSA, ECDH, ECDSA) — all of which are broken by Shor's algorithm on a fault-tolerant quantum computer. AES-128 is fine; your TLS key exchange is not. Redirect PQC migration effort accordingly: ECDHE → ML-KEM, RSA signatures → ML-DSA, keep AES where it is.

SourcesArs Technica

---

No significant new tool releases or major automation framework announcements today. Yesterday's Ansible Jinja2 regression analysis (72-hour cooldown) and Itential agent orchestration coverage were the week's primary automation stories. Next developments to watch: Nornir ecosystem releases and NANOG 97 CFP (open through April 27) with NEMOPS as focus area.

---

Gartner Raises IT Spending Forecast Amid Global Energy Crisis — AI/Cloud Decoupled

The Register reports that Gartner raised its worldwide IT spending growth forecast to 10.8% for 2026 (totaling $6.15 trillion), the day after the International Energy Agency declared the US/Israel/Iran conflict is creating the worst energy crisis the world has ever faced. The decoupling of enterprise tech spending from macroeconomic shocks is the notable signal.

Key Details:

• Data center spending up 31.7% to over $650 billion in 2026
• Server spending up 36.9% year-over-year
• AI spending totals $2.52 trillion in 2026, a 44% year-over-year increase
• Generative AI model spending growing at 80.8%
• The forecast was raised from the October 2025 projection of 9.8% growth — the bump reflects AI infrastructure buildout momentum outweighing energy/geopolitical uncertainty

So What?: This spending trajectory is self-reinforcing: infrastructure buildout → more AI services → more demand for infrastructure. For network engineers, the implication is sustained multi-year demand for AI fabric design, automation, and high-density datacenter networking skills. The job market for people who understand RoCEv2 fabrics and SONiC automation is not softening.

SourcesThe Register | Gartner

Claude Mythos — AI Cybersecurity Frontier

Covered in Top 3 above

---

Data Center World 2026: "Build for Legacy, Not Just Capacity"

Data Center World 2026 opened April 20-23 in Washington DC with the "Innovation at Scale" theme. The headline message from AFCOM executive chair Bill Kleyman: data centers now face projections of up to 17% of total US electricity consumption by 2030, and the industry needs to build infrastructure that outlasts individual technology generations — not just the current AI GPU wave.

Key themes from the conference:

• Sustainability and legacy design as strategic imperatives, not compliance boxes
• Power density planning for AI workloads that exceed traditional assumptions (50+ kW/rack common, 100+ kW/rack emerging)
• Siting strategy shifting away from traditional tier-one markets constrained by power and permits
• 500 solution providers registered — a 20%+ year-over-year growth, reflecting sustained investor and vendor momentum

So What?: "Build for legacy" is the counter-narrative to the current rush-to-AI-capacity framing. Operators who over-specialize for 2026 GPU workloads without thermal and power headroom for next-generation silicon are going to face expensive retrofits. The lesson from previous compute transitions (blade servers → hyperconverged → GPU) is that the next form factor always consumes more power than you planned for.

SourcesData Center Knowledge | Data Center World

ECL FlexGrid Santa Clara: Hydrogen + Grid Hybrid for High-Density AI

ECL announced CSC-1, a 35MW data center in Santa Clara combining on-grid electricity, hydrogen power, and natural gas through its proprietary FlexGrid architecture. The facility targets sub-1.15 PUE, with hydrogen by-product water supplying the cooling loop — eliminating freshwater draw in drought-constrained California.

Why it matters architecturally: This is the most technically interesting power strategy since Oracle's Bloom Energy fuel cell deal (covered April 15). Where Oracle is displacing grid entirely with solid oxide fuel cells, ECL's FlexGrid is a multi-source orchestration approach — using hydrogen when it's economically and environmentally optimal, grid when it's available, and managing the interoperability between them. The cooling loop innovation (using hydrogen generation by-product water) solves two constraints simultaneously: power and water.

SourcesDataCenter Dynamics | ECL / Morningstar

---

QuEra/Harvard/MIT Teraquop Milestone

Covered in Top 3 above

---

AES-128 Post-Quantum Status

Covered in Networking section above

Mythos and the Security Architecture Shift

The Claude Mythos result has a direct security architecture implication beyond the headlines: autonomous multi-step vulnerability chaining means that penetration testing — as a periodic, human-driven activity — is an increasingly inadequate security posture. The question is no longer "did we pen test last quarter?" but "do we have continuous AI-assisted vulnerability scanning running on our attack surface?" Project Glasswing is the first large-scale attempt to answer that question at infrastructure level, but the pattern needs to extend to enterprise networks.

---

• GitHub Copilot plan changes (April 22): GitHub changed individual Copilot plan pricing/features on the same day Anthropic briefly removed Claude Code from Pro plans (then reversed). Both moves suggest the AI coding assistant market is entering a re-pricing phase as providers seek sustainable unit economics. If you're evaluating AI coding tool procurement, expect more volatility in pricing over the next two quarters.

• Meta employee surveillance (The Register, April 22): Meta is reportedly installing keylogging/screenshot software on employee work PCs to build training data for AI. The irony of a privacy company surveilling its own employees is noted. The architecture lesson: if your employer issues managed devices, assume everything on them is potentially logged. Use unmanaged devices for personal communications.

• Stellantis migrating 60% of datacenter footprint to Azure (DataCenter Dynamics, April 21): Automaker plans 60% reduction in on-premises DC footprint via five-year Azure agreement. Trend signal: legacy automotive and manufacturing companies are executing the cloud migrations they deferred during 2022-2024 cost scrutiny. Network teams at these organizations are now managing hybrid connectivity at scale.

---

• NANOG 97 CFP closes April 27 — NEMOPS (Next Era of Network Management Operations, AI-assisted closed-loop management) is a focus area. If you're doing interesting work in this space, submit.
• QuEra Teraquop result will likely trigger a follow-on wave of qLDPC code implementations across trapped-ion and photonic platforms. Watch for IBM and Google responses.
• Project Glasswing is in preview — check anthropic.com/glasswing for application status if you maintain critical infrastructure code.

---

Amaze Networks Morning Briefing — published under Beeston Labs at beestonlabs.dev Pipeline: 6 domains researched | 7 primary stories | 3 quick takes | 0 dedup rejections