ATHENA

← all briefs

№ 65

Tuesday, July 28, 2026

AI & Tech Brief — July 28, 2026

AI & Tech Brief — July 28, 2026

TL;DR

  • The Model Context Protocol shipped its 2026-07-28 spec today, rebuilding MCP around a stateless request/response core so servers can sit behind plain load balancers — the biggest protocol change since remote MCP launched.
  • Anthropic revealed that Claude Mythos Preview autonomously discovered real cryptographic attacks — halving the effective key strength of NIST post-quantum candidate HAWK and beating the best-known reduced-round AES attack by 200–800×.
  • Moonshot AI released the full weights of Kimi K3, a 2.8-trillion-parameter model and the largest open-weights release ever, landing the same day Dario Amodei published Anthropic’s formal position against banning open-weights models.

Key Stories

  • MCP 2026-07-28 Spec Goes Stateless The Model Context Protocol’s newest specification, released today, retires the initialize handshake and Mcp-Session-Id header in favor of a fully stateless request/response core — every request is self-describing and can land on any instance behind a round-robin load balancer. Server-initiated flows like sampling and elicitation are redesigned as Multi Round-Trip Requests (MRTR), list responses become cacheable, method names move into HTTP headers for gateway routing, and Dynamic Client Registration is formally deprecated in favor of client metadata documents. TypeScript, Python, Go, and C# SDKs are updated today, with AWS, Cloudflare, Microsoft, and Figma among the launch ecosystem. With Tier-1 SDKs nearing half a billion downloads a month, this is the protocol’s maturation moment. Source: https://blog.modelcontextprotocol.io/posts/2026-07-28/

  • Claude Discovers New Cryptographic Attacks Anthropic researchers working with Claude Mythos Preview produced two substantial cryptanalysis results: an attack on HAWK (a third-round NIST post-quantum signature candidate) that cuts its effective key size in half — found in ~60 hours after two years of expert human review — and a meet-in-the-middle attack on 7-round AES that is 200–800× faster than the prior best, via a model-invented technique it named a “Möbius Bridge.” Neither result affects production systems, but each cost only ~$100K in API spend and was largely autonomous, signaling that frontier models now do expert-level cryptographic research. Anthropic also released CryptanalysisBench with ETH Zurich, Tel Aviv University, and University of Haifa. Source: https://www.anthropic.com/research/discovering-cryptographic-weaknesses

  • Kimi K3 Weights Released — Largest Open Model Ever Moonshot AI published the full weights and technical paper for Kimi K3, making the 2.8T-parameter frontier model free to download for research, personal, and commercial use (with guardrails). K3 debuted eleven days ago and demand was high enough that Moonshot temporarily paused new subscriptions. The release intensifies pricing and performance pressure on US frontier labs, and lands in the middle of a heated US policy debate over Chinese open-weights models. Source: https://huggingface.co/moonshotai/Kimi-K3

  • Dario Amodei: Anthropic’s Position on Open-Weights Models In a July 27 statement, the Anthropic CEO responded to reports that US officials are considering banning Chinese open-weights models — and to accusations that Anthropic backs such a ban. Amodei states plainly that Anthropic has never advocated banning open-weights models, calling non-dangerous open models a public good. Instead he argues for three targeted measures: keeping advanced chips out of China, cracking down on industrial-scale distillation operations, and mandatory safety testing for all sufficiently capable models, open or closed. Notably, Anthropic is the only major lab that did not sign the industry letter in support of open weights. Source: https://www.anthropic.com/news/position-open-weights-models

  • OpenAI Field Report: Agentic AI in Scientific Computing OpenAI published an exploratory report covering eight agent-assisted scientific software projects (five using Codex alone, three combining Codex and Claude Code), spanning genomic data libraries, language migrations, and GPU-native redesigns. The recurring finding: agents collapse the cost of implementation, shifting the researcher’s role from writing code to specifying, verifying, and stewarding it — but agents confidently produce scientifically invalid output, so external acceptance targets and long-term maintenance ownership remain the real bottlenecks. Source: https://openai.com/index/scientific-computing-agentic-ai/

  • How DoorDash, Instacart, and Uber Eats Put LLMs in Search ByteByteGo’s latest deep dive compares three production architectures for LLM-powered food search. DoorDash keeps the LLM mostly offline, enriching its knowledge graph and constraining query parsing to existing taxonomy (a clever RAG inversion, +30% carousel trigger rate). Instacart consolidated a zoo of query-understanding models into RAG-cached head queries plus a fine-tuned Llama-3-8B serving the tail in under 300ms (rewrite coverage 50% → 95%). Uber Eats went deepest, fine-tuning Qwen as the embedding backbone of its entire two-tower retrieval system across every vertical and language. The lesson: integration depth is dictated by the infrastructure you already have, not the model you pick. Source: https://blog.bytebytego.com/p/why-doordash-instacart-and-uber-eats

  • Kimi Delta Attention, Demystified The top story on Hacker News today (233 points) is a Doubleword walkthrough that derives Kimi Delta Attention — the linear-attention mechanism inside Kimi K3 — from first principles: softmax attention → linear attention → DeltaNet → Gated DeltaNet → KDA. The core trick is promoting Gated DeltaNet’s scalar retention gate to a per-channel diagonal matrix, giving a diagonal-plus-low-rank state transition that forgets each key channel independently. It includes the five-line recurrent reference implementation and the chunkwise Triton schedule used for training. Source: https://blog.doubleword.ai/you-could-have-come-up-with-kimi-delta-attention

Quiet but Interesting

  • Raschka’s Kimi K3 Architecture Notes Sebastian Raschka’s teardown of the K3 architecture: it’s essentially Kimi Linear scaled from 48B to 2.8T, plus a LatentMoE borrowed from the Nemotron 3 Ultra playbook. Notable choices — NoPE (no positional embeddings) everywhere, a first for a frontier-level model; attention residuals that connect residual paths across layers at ~4% training overhead; and native multimodality. The overall design trend is inference efficiency above all. Source: https://sebastianraschka.com/blog/2026/kimi-k3-architecture-notes.html

  • Nvidia Backs SSI; Altman Heads to Washington Per Superhuman AI’s morning edition: Nvidia is investing up to $5B in Ilya Sutskever’s Safe Superintelligence after the lab confirmed it has research “worth scaling,” and is heading a new Open Secure AI Alliance promoting open-source AI as a cybersecurity asset. Separately, Sam Altman is meeting senior US officials this week to preview OpenAI’s most powerful model, ahead of the administration finalizing its AI regulatory stance. Source: https://www.superhuman.ai/p/largest-open-weight-model-ever-is-now-available

Skip