Daily digest

15 items · ~15 min · Week 2026-W31

Must-read (3)

Anthropic's Claude Mythos Preview breaks post-quantum HAWK scheme and improves AES attack

Anthropic
Research official + media 2 src. ~1 min

Anthropic published research showing Claude Mythos Preview found a previously unknown attack that halves HAWK's post-quantum signature security margin, and a 200-800x faster attack on round-reduced AES, discovered in about 60 hours using a multi-agent research system. Anthropic also released CryptanalysisBench, built with ETH Zurich, Tel Aviv University, University of Haifa, and TU Berlin.

Why it matters
First public demonstration of AI-assisted cryptanalysis breaking meaningful security margins on schemes that survived years of expert human review, though Anthropic says no production systems are currently affected.

Google DeepMind unveils Gemini Robotics 2 for whole-body humanoid control

Google DeepMind
Research official + media 2 src. ~1 min

Google DeepMind launched Gemini Robotics 2, Gemini Robotics ER 2, and Gemini Robotics On-Device 2, giving humanoid robots whole-body reasoning so they can walk, crouch, manipulate objects, and coordinate with other robots on multi-step tasks. ER 2 is available via Google Cloud, the Gemini API, and Google AI Studio.

Why it matters
Extends Gemini's reasoning capabilities from a robot's upper body to full-body control, a capability gap competitors have not closed.

TurboVLA: Real-Time Vision-Language-Action Model at 32 Hz on an RTX 4090 with <1 GB VRAM

Research official + media 2 src. ~1 min

Reformulates vision-language-action modeling as a direct V+L to action mapping instead of routing through a large language-model backbone, reaching 97.7% average success on LIBERO with only 0.2B parameters, 31.2ms latency, and under 1GB VRAM on a consumer RTX 4090.

Why it matters
Reached 122 upvotes on HuggingFace Daily Papers for 2026-07-30; matches or beats much larger LLM-centric VLA systems while making real-time robot control feasible on commodity hardware.

Worth knowing (5)

OpenAI slashes GPT-5.6 Luna and Terra API prices, adds Fast mode

OpenAI
Industry official + media 2 src. ~1 min

OpenAI cut GPT-5.6 Luna pricing by 80% (to $0.20/$1.20 per million input/output tokens) and Terra by 20%, crediting efficiency gains from internal model-assisted optimization work. It also introduced a Fast mode for Sol in the API, up to 2.5x faster at double the price, replacing Priority Processing.

Why it matters
Signals falling frontier-adjacent inference costs and a new speed/price tradeoff lever for developers building on the API.

Metis: Memory Foundation Model

MemTensor
Research official + media 2 src. ~1 min

First prototype of a memory foundation model that gives a backbone a persistent, dynamically evolving memory state, updated gradient-free through a single forward pass rather than external retrieval or fine-tuning.

Why it matters
Ranked #1 on HuggingFace Daily Papers for 2026-07-31 with 62 upvotes; proposes memory as a native model capability rather than a bolt-on retrieval system, though the authors note it still degrades on long-horizon tasks due to fixed-size compression.

Mage-VL: An Efficient Codec-Native Streaming Multimodal Foundation Model

Microsoft
Research official + media 2 src. ~1 min

A 4B-parameter codec-native multimodal model whose custom visual tokenizer (Mage-ViT) selectively encodes motion- and residual-rich regions of video streams instead of uniform frame sampling, cutting visual token count by over 75% and giving up to 3.5x faster inference.

Why it matters
At 4B parameters it matches Qwen3-VL-4B on general VQA and surpasses the much larger Phi-4-R-V-15B on spatial and video understanding, addressing real-time perception as a known weak spot for VLMs.

Cursor launches iPad support with full PR review and split-screen chats

Cursor
Tools official 1 src. ~1 min

Cursor rolled out iPad support across all paid plans (2026-07-29), adding an inbox, full PR review surface with comments/checks/approvals, Apple Pencil markup for image annotation, Bitbucket and Azure DevOps source-control integration, multi-PR session handling, and pinned sidebar chats for monitoring multiple agents with split-screen review.

Why it matters
Extends Cursor's async-agent workflow to mobile/tablet, positioning code review and PR management as an on-the-go task rather than desktop-only.

MiniMax releases Hailuo H3 video model with native audio and open weights

MiniMax
Video official + media 2 src. ~1 min

MiniMax launched Hailuo H3 (MiniMax H3), the third generation of its Hailuo video line, generating clips up to 15 seconds at native 2K resolution with synchronized stereo audio. It supports omni-reference input (up to 9 reference images, 3 reference video clips, 3 reference audio clips), instruction-based editing, and motion transfer between videos. MiniMax said it would open-weight the model within days and priced it at roughly one-third of rival products, positioning it directly against ByteDance's Seedance 2.0 and Kuaishou's Kling 3.0.

Why it matters
A frontier-tier Chinese video model claiming to match closed competitors on quality while undercutting them on price and committing to open weights intensifies the competitive race in Chinese video generation, alongside Kling and Seedance.
For reference (7)

xAI launches Grok Voice Think Fast 2.0 voice model

xAI
Audio official + media 2 src. ~1 min

xAI released Grok Voice Think Fast 2.0, cutting reasoning-token usage by roughly 60% and time-to-first-audio from 1.25s to 0.70s versus version 1.0, priced at $0.08/minute. The default grok-voice-latest alias moves to 2.0 on August 5.

Why it matters
A notable latency and cost improvement for real-time voice AI, an area xAI is pushing to compete with OpenAI and Google's voice offerings.

Meta raises 2026 AI capex guidance to $130-145B as free cash flow collapses

Meta
Industry official + media 2 src. ~1 min

Meta's Q2 2026 results showed revenue of $60.8B (+28% YoY) but free cash flow shrank to $784M after $31.1B in AI infrastructure capex; the company raised full-year capex guidance to $130B-$145B. Reality Labs lost $4.62B in the quarter.

Why it matters
Underscores the scale of capital Meta is committing to AI infrastructure and the near-term profitability tradeoff investors are reacting to.

Oracle expands partnership with Google Cloud to bring Gemini models to enterprise apps

Google DeepMind
Industry official + media 2 src. ~1 min

Oracle and Google Cloud expanded their partnership to make Gemini 3.1 Flash-Lite and Gemini 3.5 Flash available inside Oracle AI Agent Studio for Fusion Applications and embedded in Oracle Fusion Cloud Applications and NetSuite.

Why it matters
Extends Gemini's enterprise distribution through one of the largest business-applications vendors, a distribution channel Google didn't previously have at this scale.

Sber publishes updated public threat model for AI systems covering 37 threats and 51 attack methods

Sber
Industry media only 3 src. ~1 min

Sber released an updated cybersecurity threat model for AI systems on its public 'Kibrariy' portal, describing 37 threats and 51 methods of realizing them across the full AI lifecycle, including risks specific to generative models, multi-agent systems, and RAG/LLM adapters.

Why it matters
A major Russian bank publishing a detailed, open threat taxonomy for generative and agentic AI systems gives local AI developers and security teams a concrete reference framework, reflecting growing formalization of AI security practice in the Russian market.

Codex CLI 0.146.0 adds session naming, plugin manifests, WebSocket remote connectivity

OpenAI
Tools official 1 src. ~1 min

Codex CLI 0.146.0 (released 2026-07-29) adds session naming via /new and /clear, thread pinning and paginated thread forking, Agent Plugins manifest support with Amazon Bedrock and Claude Code marketplace compatibility, WebSocket connectivity between app-server and remote Code Mode hosts, and standalone web search for custom model providers.

Why it matters
Plugin manifest interoperability with Claude Code and Bedrock signals Codex CLI moving toward a shared agent-plugin ecosystem rather than a closed format.

OpenCode v1.18.10 adds automatic Modal model discovery, desktop stability fixes

SST
Tools official 1 src. ~1 min

OpenCode v1.18.10 (released 2026-07-30) adds automatic discovery of available Modal models, prevents duplicate attachments in the desktop app, improves toast notification stacking, and fixes a bug where malformed saved tabs caused broken tab state to persist across launches.

Why it matters
Continues OpenCode's rapid weekly release cadence as it expands provider coverage (Modal) alongside desktop-app reliability work.

GitHub Copilot in Visual Studio adds SDK-based agent preview and inline code review

GitHub
Tools official 1 src. ~1 min

GitHub's July 2026 Visual Studio Copilot update (published 2026-07-30) introduces a new Copilot SDK-based agent in public preview with fewer back-and-forth turns, built-in .NET and Azure skills from Microsoft's own teams, an inline 'Review Selection' code-review feature, and organization-level custom instructions for Copilot Business/Enterprise.

Why it matters
Brings the same Copilot SDK powering GitHub Copilot CLI into Visual Studio, plus first-party .NET/Azure expertise baked into the agent's tool picker.