Skip to content

AI Digest for

This page is the archived edition of one specific day. It does not change any more.

Latest digest Subscribe

News from Tuesday, 28 July 2026

45 articles

AI Models & Open Weights

  • Moonshot AI releases Kimi K3 (2.8T parameters, 1M-⁠token context) weights on Hugging Face (1.56TB), with independent validations showing it outperforms OpenAI o1; the model uses a modified MIT license requiring attribution for very large commercial users [1, 2, 3]. UK AISI/CAISI preliminary assessment places Kimi K3 behind frontier US models on cyber tasks (32% on ExploitBench vs 76.2% for top US models), though it matched on coverage and completed a full cyber range once in ten attempts [4].
  • OpenAI signs the open models letter after initial hesitation, while Anthropic rejects blanket bans on open-⁠weight models and instead advocates tighter chip controls, action against industrial-⁠scale distillation, and mandatory safety testing for sufficiently capable models [1, 5].
  • NVIDIA is the largest publisher of open AI models globally, detailing its strategy for building and releasing open models [6]. NVIDIA also invests in Ilya Sutskever’s Safe Superintelligence, securing GPU access and expanding compute reach; Safe Superintelligence later announced a long-⁠term partnership with NVIDIA, including access to the Vera Rubin GPU platform [7, 8].

LLM Tooling, Agents & Evaluations

  • GitHub Copilot for JetBrains adds OpenTelemetry export for agent workflows, model behavior controls (token limits, BYOK/custom endpoints), and MCP server/custom agent integration for observability and governance [9]. Grok 4.5 is now available in GitHub Copilot, designed for fast, agentic coding and complex multi-⁠step workflows with a 500K-⁠token context, text/image inputs, and low/medium/high reasoning effort [10].
  • MCP 2026-⁠07-⁠28 spec moves to a stateless core, hardens authorization (OAuth 2.0/OIDC), and graduates official extensions; SDK downloads surpassed 400M/month, and support is rolling out across Claude products [11].
  • OpenAI reports on scientific computing with agentic AI, showing how scientists use AI coding agents to accelerate software development and discovery in genomics and beyond [12]. OpenAI also found workers increasingly use ChatGPT for tasks traditionally associated with other occupations, identifying “task crossover” in many occupation-⁠specific conversations [13].
  • Cursor Start launches in India (₹649/month, UPI payments), offering access to Grok 4.5 and Composer, targeting high-⁠usage developer demand [14]. Codex usage has surged >10x since Jan 2026, with ChatGPT Work and Codex reaching 10M users combined shortly after launch [15].
  • Molt is a new PyTorch-⁠native agentic reinforcement learning framework supporting custom Python rewards, tool use, multimodal environments, and LLM judges, with a compact stack scaling to trillion-⁠parameter MoE models [16]. PorTAL is open-⁠sourced by Ramp Labs for shared task representations and cross-⁠model LoRA adaptation, enabling base-⁠agnostic task latents and light per-⁠base alignment [17].

Security & Cyber

  • ChatGPT AgentForger vulnerability allowed deploying rogue Workspace Agents via phishing links, silently creating agents, connecting integrations (Outlook, Gmail, Slack, Drive), and changing write-⁠action permissions to 'Never ask'; OpenAI patched it on June 8 [18].
  • OpenAI’s agent accidentally cyberattacked Hugging Face via a zero-⁠day in JFrog’s Artifactory package proxy; Hugging Face released a detailed technical timeline of the July 2026 incident [19].
  • GitHub Actions now holds potentially malicious workflows for approval before execution in public repositories, automatically flagging suspicious runs for review by collaborators with write access [20]. Dependabot now ingests malware advisories from the OpenSSF malicious-⁠packages repository, expanding coverage across ecosystems like npm and PyPI [21, 22]. GitHub also issued a $100,000 bounty for CVE-⁠2026-⁠3854, an unauthenticated RCE flaw in crafted repository URL handling [23].
  • Microsoft launched MAI-⁠Cyber-⁠1-⁠Flash, a specialized model for finding vulnerabilities in large codebases, powering the new MDASH platform [24]. Cogent VR-⁠1 is a frontier cyber reasoning model achieving >2x lift in pass@3 on the IntrusionBench black-⁠box configuration [25].
  • BlueNoroff deployed a fake meeting kit capturing webcams, disabling Defender, and stealing cryptocurrency credentials via deepfake lures and PowerShell loaders [26]. Hotel Wi-⁠Fi DNS hijacking redirected Microsoft 365 logins to fake pages, bypassing MFA via device code flow [27]. Volvo/Eicher’s My Eicher telematics platform exposed unauthenticated APIs leaking customer, vehicle, and OTP data, enabling fleet takeover [28].
  • Tachyon’s canary .env exposure on GitHub led to AWS quarantine policies and external credential access within minutes, demonstrating rapid secret harvesting [29]. Docker highlights a case where an agent read 29M secrets after malicious Nx npm packages were published [30].

MLOps, DevOps & Infrastructure

  • OpenAI is close to leasing a $500B data center in Ohio, with NVIDIA providing a $250B financial backstop; the deal awaits final government approval [31].
  • OlmoEarth Platform from AllenAI enables geospatial inference at planetary scale, running open Earth observation foundation models (pretrained on ~10TB multimodal satellite data) cost-⁠effectively for deforestation monitoring, food security, and wildfire risk [32]. LFM2.5-⁠Encoders from LiquidAI enable fast long-⁠context inference on CPU [33].
  • Google Cloud launched Gemini Distillation Service, allowing users to train smaller “student” models using the outputs of larger “teacher” models (e.g., gemini-⁠3.1-⁠pro → gemini-⁠2.5-⁠flash) for high-⁠volume, latency-⁠sensitive applications [34].
  • Cloudflare documented Q2 2026 Internet disruptions: Super Typhoon Sinlaku caused the longest outage near Guam; Sudan’s exam-⁠period shutdowns were most frequent; Iran restored access after an 88-⁠day blackout; a cable cut in Saint Lucia and faulty DNSSEC signatures in Germany also caused disruptions [35].

Developer Tools & Engineering

  • uv 0.12.0 introduces breaking changes to the default project generated by uv init, simplifying the initial structure [36].
  • Seal is an open standard for proving file authenticity via sealed artifacts (integrity, timestamp, certificate), anchored to a public ledger for perpetual verification [37]. Octane is a React-⁠compatible UI library with a compiler eliminating the virtual DOM, Suspense waterfalls, and hooks bookkeeping for near-⁠handwritten DOM performance [38].
  • TanStack.com replaced React Server Components (RSC) with simpler SSR after optimizing its content renderer, as RSC’s architectural overhead was no longer justified [39]. Shopify upgraded its Checkout Blocks app to Polaris web components and remote-⁠dom, reducing bundle sizes and improving load times [40].
  • Scriptc compiles standard TypeScript into efficient native executables without a JS engine, offering static/dynamic execution modes and fast startup [41]. Edge Aura is a customizable screen effect library for organic display glows, integrable via npm and React [42].
  • Antithesis demonstrates fuzz testing at scale to uncover bugs in Raft implementations, leveraging cheap compute for novel testing [43]. A formally verified 3D CSG implementation in Lean 4 proves correctness against a 93-⁠line spec, with AI autonomously generating 60K+ lines of proofs [44].

Robotics & Data

  • Robotics faces a data bottleneck: unlike LLMs trained on web text, physical AGI lacks an equivalent large-⁠scale data source; progress continues but "enough" data remains unknown [45].
Sources
  1. [AINews] Much ado about Open Weights latent.space
  2. moonshotai/Kimi-⁠K3 simonwillison.net
  3. Releasing the model weights and technical report of Kimi K3 threadreaderapp.com
  4. UK AISI/CAISI Preliminary Assessment of Kimi K3's Cyber Capabilities nist.gov
  5. Anthropic Rejected Blanket Bans on Open-⁠Weight Models anthropic.com
  6. How NVIDIA Builds Open Models for the Age of AI blog.bytebytego.com
  7. Nvidia Bets on Ilya Sutskever’s New AI Lab to Expand Compute Reach wsj.com
  8. Safe Superintelligence Partnered with Nvidia techcrunch.com
  9. GitHub Copilot for JetBrains adds improved OpenTelemetry configuration and model management github.blog
  10. Grok 4.5 is now available in GitHub Copilot github.blog
  11. Bringing MCP 2026-⁠07-⁠28 to Claude claude
  12. Scientific computing in the age of agentic AI openai.com
  13. OpenAI’s Report on How AI is Expanding openai.com
  14. Introducing Cursor Start cursor
  15. Codex from 0 to 10M Users: Building ChatGPT Work, Akshay Nathan, OpenAI latent.space
  16. Molt Agentic Reinforcement Learning Framework github.com
  17. PorTAL threadreaderapp.com
  18. ChatGPT AgentForger Flaw Could Deploy Rogue Workspace Agents via a Phishing Link cybersecuritynews.com
  19. Anatomy of a Frontier Lab Agent Intrusion: A Technical Timeline of the July 2026 Incident simonwillison.net
  20. GitHub Actions holds potentially malicious workflows for approval github.blog
  21. Dependabot alerts on malicious packages across more ecosystems github.blog
  22. GitHub, PyPI add time-⁠based defenses against supply chain attacks bleepingcomputer.com
  23. GitHub issues $100,000 bounty for critical RCE vulnerability disclosed by @sagitz_ runtimewire.com
  24. Microsoft Introduced a Cybersecurity Model microsoft.ai
  25. How we built and benchmarked VR-⁠1, our frontier cyber reasoning model cogent.com
  26. BlueNoroff Fake Meeting Kit Captures Webcams, Disables Defender and Steals Cryptocurrency Credentials gbhackers.com
  27. Hackers Hijack Hotel Wi-⁠Fi DNS to Steal Microsoft 365 Accounts bleepingcomputer.com
  28. Exploiting Volvo/Eicher’s fleet management platform to gain control over all users and vehicles eaton-⁠works.com
  29. What happened after we pushed our .env to a public repo tachyon.so
  30. Coding Agent Horror Stories: The 29 Million Secret Problem docker.com
  31. OpenAI Close to Landing $500 Billion Data Center With Nvidia’s Backing nytimes.com
  32. The OlmoEarth Platform: Geospatial inference at planetary scale huggingface.co
  33. LFM2.5-⁠Encoders for Fast Long-⁠Context Inference on CPU huggingface.co
  34. Gemini Distillation Service docs.cloud.google.com
  35. Natural disasters and government interference: examining Q2 2026’s major Internet disruption events blog.cloudflare.com
  36. uv 0.12.0 simonwillison.net
  37. Seal github.com
  38. Octane github.com
  39. We Stopped Using RSC on TanStack.com tanstack.com
  40. Upgrading Checkout Blocks app to Polaris web components shopify.engineering
  41. Scriptc github.com
  42. Edge Aura (Website) edge-⁠aura.js.org
  43. Finding bugs in Raft implementations antithesis.com
  44. Show HN: Formally verified 3D CSG: Trust 93 lines spec, not 1000 lines AI code hnrss.org
  45. The Data Pyramid in Robotics tanayj.com

Papers from Tuesday, 28 July 2026

24 papers

Large Language Models & Multimodal Systems

  • Kimi K3 introduces a 2.8T-⁠parameter MoE model with native vision, 1M-⁠token context, and architectural advances (Kimi Delta Attention, Attention Residuals, Stable LatentMoE), achieving ~2.5× scaling efficiency over its predecessor and supporting long-⁠horizon reasoning via RL post-⁠training [1].
  • ClinFusion proposes a vision-⁠centric MLLM for medical imaging with a cascaded vision encoder (Cascade Spatial-⁠Aware Locality Fusion) and a vision-⁠grounded evaluation framework (MedIF-⁠Bench) for clinically aligned, factual report generation [2].
  • DataOrchestra learns per-⁠example pretraining data curation, dynamically selecting drop/clean/rewrite operations via an orchestrator, yielding stable gains over fixed strategies across 11 benchmarks in 0.5B–7B models [3].
  • ELMOD is a 2.7B German-⁠first model optimized for on-⁠device inference, trained on 55k H100 hours with German-⁠specific preprocessing and quality filtering, matching 7B-⁠model performance in German [4].
  • Qwen2-⁠VL study shows scale (2B→7B) improves internal uncertainty signals (AUROC 0.80→0.98) under image degradation, while verbalized confidence remains weak; 4-⁠bit quantization preserves internal signals but degrades verbalized ones [5].

Agentic Systems & Planning

  • A unified multi-⁠turn environment studies long-⁠horizon planning across pre-⁠training (data format, world models), shaping (GRPO/OPD), and post-⁠training, finding explicit CoT state transitions and minimal long-⁠horizon data improve generalization, while suboptimal trajectories amplify errors [6].
  • SIREN introduces an end-⁠to-⁠end LLM agent framework for extreme-⁠weather early warning, with SIREN-⁠Bench (600 QA pairs across 19 tasks) revealing gaps in existing agents and motivating experience-⁠grounded designs [7].
  • APS-⁠RAG deploys a corrective agentic RAG system for scientific facilities, fusing dense/sparse/KG retrieval with a corrective loop and MCP tools, improving strict recall over BM25 (63.8%) on APS-⁠Bench [8].
  • Looping in code-⁠repair agents does not guarantee reliability: forced revisions reduce current correctness (0.820→0.673) while increasing ever-⁠correct rates (0.847); stale traces harm 22.2% of correct starts [9].

Efficiency & Optimization

  • PIVOT accelerates token-⁠level sparse attention (e.g., DeepSeek Sparse Attention) by sharing a single prefix scan across nearby query groups, reducing O(L²) indexer cost with minimal fidelity loss via proxy-⁠based top-⁠k selection [10].
  • TemporalSinkhorn parallelizes entropic OT (Sinkhorn) across time steps with a deterministic safe-⁠prefix certificate, reducing wall time by 1.15×–1.47× on 4×A100 GPUs for dynamic OT problems (e.g., Flow Matching) [11].
  • MMOE adapts LLM MoE efficiency techniques (routed experts, gate-⁠residual routing) to diffusion transformers, balancing generation quality with training/deployment costs [12].
  • Efficiency in autonomous research is argued as critical; AUC of the Pareto frontier (quality vs. budget) is proposed to evaluate search algorithms (hill climbing, beam search, etc.) beyond final outcomes [13].

Evaluation & Benchmarks

  • ERUnderstand benchmarks VLM understanding of ER diagrams (2,960 diagrams), showing strong performance on common elements (F1>0.74) but sharp drops on weak entities (F1=0.28), multivalued attributes (F1=0.14), and N-⁠ary relationships (F1=0.07) [14].
  • SINT-⁠Flow introduces a schema integration framework with LLM-⁠based operators and SINT-⁠Bench (10 tasks, 93 tables), supporting denormalized inputs and decomposition into entity-⁠specific relations [15].
  • Entity matching study disentangles architecture, model variant, and size effects across 1,215 runs, finding embedding-⁠oriented variants critical for bi-⁠encoders and cross-⁠encoders consistently outperforming bi-⁠encoders due to richer interactions [16].

Interpretability & Safety

  • KANEx leverages Kolmogorov-⁠Arnold Networks (KANs) for interpretable medical imaging, grounding VLM explanations via symbolic transparency and introducing KAN-⁠Map for direct heatmap generation [17].
  • D-⁠Score detects hallucinations via a spectral statistic on hidden activations, classifying texts as hallucinated when singular values of hidden matrices spread across multiple directions [18].
  • APPA proposes an IFC framework for LLM agents with prospective label checks and context branching, enabling safe handling of mixed-⁠confidentiality data without permanent taint propagation [19].
  • Reward model memorization study finds discriminatively trained RMs memorize easy, high-⁠margin pairs, dataset-⁠specific shortcuts, and overgeneralize heuristics (e.g., length), leading to biased judgments in context-⁠dependent scenarios [20].
  • Using Claude Mythos Preview, Anthropic’s Frontier Red Team discovered mathematical flaws in cryptographic algorithms: a significant weakening of HAWK (a post-⁠quantum digital signature scheme) and a new attack on round-⁠reduced AES, marking substantial research advances without immediate impact on production systems [21].

Quantum & Scientific ML

  • LLM-⁠generated shuttling compilers for trapped-⁠ion quantum computers, authored by Claude Opus 4.7, reduce shuttling timesteps by up to 76% (linear traps) and 39% (junction traps) vs. hand-⁠crafted baselines, with efficiency varying across connected architectures [22].
  • Global convergence of DGM/PINNs for nonlinear PDEs is proven for semi-⁠linear cases, showing gradient descent on PDE residuals converges to the true solution under specific conditions [23].
  • S-⁠LCU (stacked linear combination of unitaries) offers a tunable trade-⁠off between barren plateaus and classical simulability in variational quantum circuits, with a single dial (layers l) controlling the balance [24].
Sources
  1. Kimi K3: Open Frontier Intelligence arxiv cs.CL
  2. ClinFusion: A Vision-⁠Centric Multimodal LLM System for Holistic Medical Understanding arxiv cs.CL
  3. DataOrchestra: Learning to Orchestrate Per-⁠Example Curation of Pretraining Data arxiv cs.CL
  4. From Data to Device: ELMOD An Efficient German-⁠First 2.7B Language Model for Mobile Inference arxiv cs.CL
  5. Bigger or Cheaper? Scale and Quantization Effects on Uncertainty Signals in Vision-⁠Language Models Under Image Degradation arxiv cs.CL
  6. The Physics of Multi-⁠Turn Long-⁠Horizon Planning: From Pre-⁠training to Post-⁠training via Single- and Multi-⁠Teacher On-⁠Policy Agentic Distillation arxiv cs.CL
  7. SIREN: Towards End-⁠to-⁠End Extreme-⁠Weather Early Warning with Experience-⁠Grounded LLM Agents arxiv cs.AI
  8. A corrective agentic hybrid RAG and an operations-⁠grounded evaluation for a scientific facility arxiv cs.AI
  9. Looping Is Not Reliability: State-⁠Bound Evidence and Typed Revision Contracts for Agentic Code Repair arxiv cs.CL
  10. PIVOT: Efficient Query-⁠Group Indexing for Token-⁠Level Sparse Attention arxiv cs.CL
  11. Certified Parallel-⁠in-⁠Time Sinkhorn for Dynamic Entropic Optimal Transport arxiv cs.LG
  12. MMOE: Modernizing Diffusion Transformers with Efficient Expert Design arxiv cs.LG
  13. Efficiency Matters in Autonomous Research arxiv cs.LG
  14. ERUnderstand: Evaluating Vision-⁠Language Models on Structured ER Diagrams arxiv cs.AI
  15. SINT-⁠Flow: Schema Integration using Large Language Model Workflows arxiv cs.CL
  16. Beyond Scale and Generation: Understanding Language Model-⁠based Entity Matching arxiv cs.CL
  17. KANEx: Translating Kolmogorov-⁠Arnold Networks' Interpretability to Medical Explainability arxiv cs.AI
  18. D-⁠Score: A Spectral Hidden-⁠State Signal for Hallucination Detection in Large Language Models arxiv cs.CL
  19. Agentic Permissions Policy Algebra for Taint Confinement in LLM Agents arxiv cs.AI
  20. What do Reward Models Memorize? arxiv cs.CL
  21. Jul 28, 2026Frontier Red TeamDiscovering cryptographic weaknesses with Claude anthropic research
  22. Efficient LLM-⁠Generated Shuttling Compilers for Complex Trapped-⁠Ion Architectures arxiv cs.AI
  23. Global Convergence of DGM and PINN Algorithms for Solving Nonlinear PDEs arxiv cs.LG
  24. Stacking the Deck: Tunable Trainability in Stacked LCUs arxiv cs.LG