MDASH Tops Mythos on Cyber, SubQ Goes Sub-Quadratic, A2A Lands Against MCP, Architecture Diversifies: Sunday Digest, May 17, 2026

The single most consequential capability datapoint in the cycle is Microsoft’s MDASH— a multi-agent vulnerability-scanning system that orchestrates more than one hundred specialized agents across multiple models — scoring 88.45% on CyberGym and beating Anthropic’s Mythos alongside every other single-model frontier system on the benchmark. CyberGym is a recognized industry yardstick for end-to-end agentic cyber work, and the result is the cleanest production proof to date that orchestrated multi-agent systems can outscore the best single model on a narrow, high-stakes task.

The strategic read is not that Mythos is suddenly second-tier — it is that the competitive frontier on cyber is now system-shaped, not model-shaped. Mythos’s Project Glasswing preview remains the single most consequential offensive-capability datapoint we have. But MDASH’s win shows that a defender stitching together a hundred narrower agents, pulling from multiple model families, and routing tasks through a planner can beat a monolithic system inside a measured arena. The 88.45 figure is meaningful precisely because it was scored against the model the U.S. and EU policy machinery is currently reorganizing around.

Two adjacent items round out the cyber arc. OpenAI continues to open the EU door on GPT‑5.5‑Cyber via a vetted-defender preview — the same regulatory-side asymmetry we’ve flagged in prior briefings while Anthropic holds Mythos out of the EU. And the broader 2026 reality — that frontier systems are routinely surfacing previously-unknown vulnerabilities — is no longer a research projection. It is now the default assumption inside every patch-cycle planning meeting we’ve seen this month.

SubQ launched the first commercially available LLM built on a fully sub‑quadratic sparse-attention architecture. It is not a standard transformer: attention scales below O(n²), the model ships with a native 12M-token context window, and the vendor claims ~1/5 the inference cost of frontier dense models on the long-context workloads where attention dominates the bill. Independent quality bake-offs are still landing, but if the numbers hold up at parity-class reasoning, SubQ is the most credible architectural challenge to dense transformer attention shipped to a paying market.

The architectural pressure is not arriving alone. Google’s TurboQuant (ICLR 2026) keeps moving on a parallel track — KV-cache compression via PolarQuant vector rotation and a Quantized Johnson–Lindenstrauss step, attacking the same long-context cost surface from the deployment side rather than the architecture side. Cloudflare’s Unweight bundles with TurboQuant on Workers AI to reprice 100K-plus-token inference at the edge. And underneath both, disaggregated prefill/decode serving is now standard inside the largest serving stacks. The composite read: the long-context unit economics of April were the high-water mark, and the curve is bending fast.

For the open side of the market, Zyphra’s ZAYA1‑8B matters for a different reason: it is an 8B MoE (~760M active parameters, Apache 2.0) and the first frontier-class model trained end-to-end on AMD Instinct hardware — not ported, not fine-tuned. That is the single cleanest training-side credibility moment AMD has had in years and a real shot at relaxing NVIDIA’s grip on the training market. Add NVIDIA’s Nemotron 3 family(Nano/Super/Ultra, agent-tuned), and the open-weight tier now has fresh agent-specific anchors at multiple cost points.

Microsoft Agent Framework 1.0 shipped generally available for .NET and Python — a production-ready, open-source multi-agent framework — and it ships with the A2A (Agent-to-Agent) protocol for cross-runtime structured messaging. A2A is the cleanest competitor yet to Anthropic’s MCP, which crossed 97M installs and moved to Linux Foundation governance earlier this month. The interop standards layer is no longer a will-this-converge question; it is a which-protocol-wins-which-runtime question, and the answer matters for every team building agent workflows that span more than one vendor.

Two adjacent moves shape the standards math. Google’s Gemini Enterprise Agent Platform — Vertex AI rebranded and consolidated with Agentspace — ships a no-code agent builder (Workspace Studio), 200+ models in Model Garden, and Project Mariner web-browsing agents. Google has not yet picked a public side in the A2A–MCP debate. AWS has been similarly noncommittal, though AgentCore Payments and adjacent rails imply an MCP-friendly default. The short-term reality is that customers will run both protocols; the medium-term question is whether one hyperscaler’s endorsement collapses the choice for everyone else.

Elsewhere on the developer surface, the cycle kept the pressure on. Claude Code sits alone at the top of coding-agent rankings this month, with 1M-token context now GA at no premiumon the $200/mo Max tier. Cursor 3.2 shipped /multitask parallel sub-agents, PR Review, parallel Build Plan, and Split PRs. Notion’s Developer Platform (May 13) opened Workers, an External Agent API, and database sync — custom code and outside agents now run inside Notion. And Hermes Agent (Nous Research) crossed 140k+ GitHub stars in under three months and is, per OpenRouter, the most-used agent in the world right now.

The research layer kept its post-LLM pace. Anthropic’s “Dreaming”technique — agents reviewing prior sessions between runs to extract patterns and improve future performance — continues to harden into the long-running autonomous-agent story Anthropic has been building since Q1. OpenAI’s RLHF 2.0 pipeline cut alignment tax ~60% vs. prior methods, with independent red-team evaluations showing ~90% reduction in successful adversarial attacks on GPT‑5.1 relative to GPT‑5 — the cleanest single safety-side improvement we have seen out of OpenAI this year. Both techniques will land downstream as features, not papers.

On the math-meets-ML side, UPenn’s Mollifier Layers embed classical smoothing functions into neural networks to stabilize inverse-PDE solvers in noisy real-world data, with applications in genomics, climate, materials, and chromatin biology. University of Hawaiʻi Mānoa published a physics-informed ML algorithm that more rigorously enforces physical laws during training, targeted at fluid dynamics and climate. Separately, MIT News surfaced a fresh LLM training-efficiency method worth a follow-up read. The post-LLM “AI-for-science” arc that started lighting up in April is no longer occasional — it is now the steady drumbeat alongside frontier releases.

Two structural items round out the model layer. NVIDIA’s Nemotron 3 family(Nano/Super/Ultra) ships open agent-tuned weights at multiple cost points; Nano is available now and is optimized for debugging, summarization, retrieval, and assistant workflows at low inference cost. GPT‑5.5 Instant (May 5) replaced GPT‑5.3 Instant as the ChatGPT default for free and paid tiers, completing the GPT‑5.5 family’s consumer rollout. Combined with Opus 4.7 GA, the consumer-tier frontier picture for May is now stable: GPT‑5.5 Instant in ChatGPT, Opus 4.7 in Claude, Gemini whatever-Google-ships-Tuesday in Gemini.

The EU policy surface keeps shifting around the Mythos question. Anthropic continues to withhold Mythos from the European Union on cyber-capability grounds. OpenAI, by contrast, agreed earlier this month to open EU access to GPT‑5.5‑Cyber via a vetted-team preview for defenders. The asymmetry is now a real regulatory-side strategic gap: any EU buyer needing a frontier-tier cyber-capable model in 2026 has a clear path to OpenAI and an uncertain one to Anthropic.

Underneath, the EU AI Act “Omnibus” agreement (May 7) shifted the compliance calendar in both directions at once. The regulatory-sandbox deadline pushed to August 2027 while the transparency-grace-period for AI-generated content was cut from 6 to 3 months — a new effective deadline of December 2, 2026. Full AI Act applicability remains August 2, 2026. On the bilateral side, Treasury Secretary Bessent’s U.S.–China protocol talks continue to inch toward a working agreement on keeping advanced models out of nonstate-actor hands. And on the U.S. domestic side, the White House is studying a pre-release AI security executive order, with NEC director Hassett publicly comparing the proposed review to FDA drug evaluation.

Three threads for the early week. First, Google’s next Gemini drop and Workspace/Android agentic integration: any further movement on cross-app agentic tasks, contextual summarization, smart form-fill, and the rumored Omni video-generation system would extend the Android-as-agent arc and tighten the consumer-agent race ahead of Apple’s reboot. Second, Mythos’s broader rollout: confirmation of a wider Mythos GA, and whether EU access materializes, would redraw the GPT‑5.5‑Cyber asymmetry in a single news cycle. Third, A2A vs. MCP endorsements from Google and AWS: a public bless from either hyperscaler — integration announcements, default protocol selections, or marketplace tags — would settle the standards math for the long tail of enterprise tooling faster than any spec revision.