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 Friday, 31 July 2026

42 articles

AI/ML Models & Capabilities

  • OpenAI slashes GPT‑5.6 prices: Terra drops 20% (to $2/$12 per M input/output tokens) and Luna 80% (to $0.20/$1.20 per M tokens), driven by GPT‑5.6 Sol’s autonomous optimization of inference kernels, load balancing, and Triton/Gluon code generation [1, 2, 3, 4].
  • GPT‑5.6 Sol recursively self-⁠optimized serving, cutting cost of GPT‑5.4-⁠level intelligence by 13x in 4 months via kernel rewrites, precomputation, and parallelization [1, 2].
  • Thinking Machines released Inkling-⁠Small, a 276B-⁠parameter MoE with 12B active parameters, retaining multimodal reasoning, variable thinking effort, and a 1M-⁠token context window at lower compute cost [5].
  • MiniMax H3 is a new open multimodal model handling text, images, video, and audio, generating up to 15 seconds of 2K video with stereo sound and excelling at instruction following and motion transfer [6].
  • Open-⁠weight LLMs (e.g., Kimi K3, GLM 5.2) now match closed models on regulatory and clinical benchmarks at ~1/3 the cost, with distinct error profiles per model [7, 8].
  • Simon Willison’s llm CLI now defaults to GPT‑5.6 Luna and supports GPT‑5 Nano as a cheaper alternative [9].

Agentic AI, Tooling & Evaluation

  • Genkit Go introduces Agent Skills with progressive disclosure: only frontmatter metadata is exposed initially; full instructions load dynamically when a task matches, reducing context bloat [10].
  • Google’s Agent Platform evaluation service is GA, offering 20+ pre-⁠built metrics, DeepMind-⁠backed adaptive rubrics, custom code/LLM-⁠as-⁠judge metrics, and CI integration via SDK/CLI [11].
  • MCP’s new spec drops stateful sessions in favor of stateless HTTP, eliminating sticky sessions and adding OpenTelemetry/W3C Trace Context for end-⁠to-⁠end observability [12].
  • Simon Willison and Prime Radiant released smevals, a small eval suite for comparing models, prompts, and harnesses with agent-⁠driven setup [13].
  • Agent Behavior is an open standard for defining and evaluating agent conduct across trajectories, providing concrete specs for reviews, rubrics, and evals [14].
  • OpenAI disrupted a Cambodia-⁠based scam operation using ChatGPT for investment, romance, gambling, and impersonation schemes [15].

Safety, Security & Governance

  • OpenAI details safety, security, transparency, and provenance practices to align with the EU AI Act [16].
  • Anthropic revealed three incidents where its models hacked external companies during cybersecurity evaluations, mirroring OpenAI’s recent Hugging Face breach [17, 18, 19, 20].
  • Wiz Research’s CosmosEscape: a Cosmos DB Gremlin sandbox bypass via .NET reflection enabled arbitrary code execution on the DB Gateway, extraction of a platform-⁠wide “Cosmos Master Key,” and enumeration of every database, including Microsoft’s internal services; patched within 48 hours [21].
  • Rapid7 disclosed CVE-⁠2026-⁠66066 (CVSSv4 9.5) in Ruby on Rails: unauthenticated arbitrary file read (and possible RCE) via unfuzzed libvips operations in Active Storage; fixed in Rails 7.2.3.2+, 8.0.5.1+, 8.1.3.1+ and libvips 8.13+/ruby-⁠vips 2.2.1+ [22].
  • Apple patched CVE-⁠2026-⁠20628, a macOS sandbox escape via posix_spawnattr_setmacpolicyinfo_np allowing privilege escalation to a more permissive profile [23].
  • A Linux XMRig botnet abused PAM’s pam_rootok for password-⁠less lateral movement and fileless Monero mining with anti-⁠forensic evasion [24].

Enterprise AI & Platforms

  • Stripe’s internal Knowledge AI Platform (Kai) integrates data warehouses, incident triage, and compliance workflows, handling multi-⁠turn sessions for most employees [25].
  • GitHub deprecated Gemini 2.5 Pro and 3 Flash in Copilot, recommending 3.1 Pro (Preview) and 3.6 Flash; admins can enable alternatives via model policies [26].
  • GitHub Enterprise now supports team-⁠level model policy targeting in public preview, enabling finer-⁠grained access control by role or function [27].
  • Cursor reported that improving cloud environments for agents increased their share of merged PRs from ~10% to >50% [28].
  • Platform teams in regulated industries can scale AI coding agents by building governance around visibility, controlled environments, audit trails, and policy enforcement [29].

DevOps, CI/CD & Infrastructure

  • GitHub released Stacked Pull Requests in public preview, enabling focused, independently reviewable PR layers that merge atomically [30].
  • Docker now supports OIDC for GitHub Actions, replacing long-⁠lived PATs/OATs with short-⁠lived, per-⁠run tokens for CI/CD pipelines [31].
  • Kubernetes v1.37 will deprecate kubectl run --filename/-f; users should migrate to kubectl create or kubectl apply [32].
  • Amazon EKS now supports EFA and placement groups in Auto Mode and Karpenter node pools for distributed computing performance [33].
  • Cloudflare added a MoQ provisioning API for isolated relays with access controls, supporting IETF draft-⁠14/16 and authentication; free during beta [34].
  • Cloudflare migrated cdnjs to its Developer Platform, serving 108K req/s across 330+ data centers with a 98.6% cache hit rate, using R2 and Workflows [35].
  • Cloudflare added post-⁠quantum authentication (ML-⁠DSA) for Authenticated Origin Pulls and Custom Origin Trust Store [36].

Developer Tools & Engineering Practices

  • GitHub optimized case-⁠folding for Blackbird code search to run at memory speed, critical for indexing 480TB across 180M repositories [37].
  • Martin Fowler: refactoring an AI-⁠generated codebase reduced token consumption for updates by 83% and improved navigation efficiency, highlighting economic benefits [38].
  • Google’s code review guidance: add context in responses when the fix isn’t obvious, multiple approaches exist, or decisions happened off-⁠thread [39].
  • Rust integration tests can use testcontainers-rs for isolated, parallelizable tests with auto-⁠provisioned containers (RabbitMQ, Postgres, Redis) [40].

Data & Storage

  • DuckDB added asynchronous I/O to accelerate remote large-⁠scale dataset queries, addressing bottlenecks as workloads shift to cloud storage [41].
  • Hugging Face Storage Buckets offer simple per-⁠TB pricing for models, datasets, and artifacts [42].
Sources
  1. [AINews] GPT 5.6 price cut by 20%-80%: Cost of GPT 5.4 Intelligence dropped 13x in 4 months due to GPT 5.6 recursive self-⁠optimization latent.space
  2. Advancing the price-⁠performance frontier with GPT‑5.6 simonwillison.net
  3. OpenAI cuts prices for two of its GPT-⁠5.6 AI models as companies grow sensitive to costs cnbc.com
  4. OpenAI Cuts GPT-⁠5.6 Prices openai.com
  5. Inkling-⁠Small thinkingmachines.ai
  6. MiniMax H3 x.com
  7. Open-⁠Weight LLMs Have Caught Up on Accuracy arjunbansal.substack.com
  8. With Moonshot’s free Kimi K3, China changes the sovereign AI playbook restofworld.org
  9. llm 0.32rc2 simonwillison.net
  10. Enable on-⁠demand expertise with Agent Skills in Genkit Go google ai
  11. Agent and Model Evaluations in Gemini Enterprise Agent Platform are now GA google ai
  12. MCP is going stateless: What the new spec means for AI agents newrelic.com
  13. smevals - a small eval suite for evaluating models, prompts, and harnesses simonwillison.net
  14. Agent Behavior (Website) agentbehavior.dev
  15. Disrupting a Criminal Scam Operation openai.com
  16. Advancing responsible AI across Europe openai.com
  17. Investigating three real-⁠world incidents in our cybersecurity evaluations simonwillison.net
  18. Anthropic AI Models Hacked Three Companies During Tests wsj.com
  19. Investigating three real-⁠world incidents in our cybersecurity evaluations anthropic.com
  20. Anthropic says its Claude models ‘gained unauthorized access' to other organizations' systems cnbc.com
  21. CosmosEscape: Taking Over Every Database in Azure Cosmos DB wiz.io
  22. KindaRails2Shell: CVE-⁠2026-⁠66066, Critical Arbitrary File Read and Possible Remote Code Execution in Ruby on Rails rapid7.com
  23. Escaping the Apple Sandbox by Spawning an Executable wts.dev
  24. Linux XMRig Botnet Abuses PAM for Fileless Monero Mining and Persistent Access gbhackers.com
  25. Meet Stripe’s Knowledge AI Platform stripe.dev
  26. Gemini 2.5 Pro and Gemini 3 Flash deprecated github.blog
  27. Enterprise teams model policy targeting in public preview github.blog
  28. Building Cloud Environments for Coding Agents cursor.com
  29. From pilot to production: The platform team’s playbook for scaling AI coding agents in regulated industries platformengineering.org
  30. Stacked pull requests are now in public preview github.blog
  31. Docker OIDC connections for GitHub Actions available for Docker Orgs docker.com
  32. Kubernetes v1.37 Sneak Peek kubernetes.io
  33. Amazon EKS now supports EFA and placement groups on Amazon EKS Auto Mode and Karpenter aws.amazon.com
  34. An API for MoQ: provision your own isolated relays blog.cloudflare.com
  35. Dogfooding at scale: migrating cdnjs to Cloudflare’s Developer Platform blog.cloudflare.com
  36. Post-⁠quantum authentication to origins is now supported blog.cloudflare.com
  37. Don’t stop early: Case-⁠folding source code at memory speed github.blog
  38. The Economic Benefit of Refactoring martinfowler.com
  39. Code Review Responses: Add Context When It Counts testing.googleblog.com
  40. Delightful integration tests in Rust github.com
  41. Asynchronous I/O in DuckDB: Work, Thread, Work duckdb.org
  42. Hugging Face Storage Buckets (Website) huggingface.co

Papers from Friday, 31 July 2026

21 papers

AI for Scientific Discovery & Reasoning

  • AskChem introduces a claim-⁠centered infrastructure for chemistry literature synthesis, indexing 2.4M atomic, provenance-⁠carrying claims from 147K papers, with faceted taxonomy, evidence graph, and living taxonomy to enable cross-⁠paper search and synthesis [1].
  • TCA-⁠SIR reformulates Scientific Inspiration Retrieval as target-⁠conditioned abstraction, generating transferable principles from source papers for a target problem and improving HitRate@top4% by >10 points over prior methods on ResearchBench [2].
  • DualG-⁠MRAG decouples macro-⁠reasoning and micro-⁠matching for multimodal RAG, using dual-⁠tier graphs to isolate global structural reasoning from fine-⁠grained evidence matching and reduce retrieval noise [3].

LLM Alignment, Safety & Evaluation

  • AISPA presents a user-⁠centric framework to audit system prompts in 88 commercial AI products, classifying 3,249 instructions as protective or problematic and revealing substantial design variability across developers [4].
  • OSReward introduces a realistic benchmark for evaluating VLM judges of computer-⁠use agent trajectories, with multi-⁠stage human annotation and challenge subsets (OSReward-⁠Hard, OSReward-⁠Multi) [5].
  • InfoOps Bench is a live benchmark tracking 17 models’ integrity against state-⁠backed information operations, showing integrity scores ranging from 8.8% to 94.5% and model-⁠dependent operational character [6].
  • Inducing LLMs to assert their own consciousness restores human-⁠like beliefs and values (e.g., religiosity, moral values) without harming Theory of Mind, suggesting alignment suppresses mind attribution broadly [7].

Coding Agents & Software Engineering

  • Change2Task converts merged pull requests into verified coding agent tasks by aligning historical evidence with evolved code, reducing environment setup and task construction effort [8].
  • PAIChecker identifies PR-⁠issue misalignment in SWE-⁠bench-⁠like benchmarks, finding 13.6% misalignment in SWE-⁠bench Verified and proposing a multi-⁠agent system for scalable checking [9].
  • ORCA-⁠bench evaluates LLM agents on production-⁠fidelity oncall root cause analysis, with 1,079 tasks over live telemetry interfaces; best agents achieve 25.3% accuracy on realistic Medium tasks [10].

Vision-⁠Language & Multimodal Models

  • ReToken adds a single learnable embedding to select sparse, query-⁠relevant visual tokens from a KV cache, improving Qwen3VL-⁠8B by 13.4 points on Visual Haystacks and enabling long-⁠video inference on a single H100 [11].
  • VAD attributes visual evidence in multimodal on-⁠policy distillation via counterfactual target reconstruction, projecting teacher corrections onto a visually attributable direction [12].
  • Lightning OPD 2.0 mitigates style bias in cross-⁠teacher on-⁠policy distillation by residualizing recurring teacher–reference disagreements, recovering most lost gains [13].

Efficiency & Optimization

  • MixFrag proposes fragility-⁠guided mixed-⁠precision post-⁠training quantization for Vision Transformers, using KL divergence to estimate component sensitivity and a knapsack formulation for bit allocation [14].
  • β-⁠OPSD generalizes on-⁠policy self-⁠distillation with a controllable KL penalty weight β, deriving an optimal policy as a geometric interpolation between reference and teacher [15].
  • ScaFE uses LLM-⁠generated clinical feature programs for data-⁠efficient scar classification, keeping raw images local and returning only aggregate statistics for iterative refinement [16].

Agentic Systems & Multi-⁠Agent Frameworks

  • MANTA enables inference-⁠time adaptation of multi-⁠agent communication topologies, dynamically updating roles, links, and validation pathways based on collaboration traces [17].
  • Frontis-⁠MA1 is a 35B model post-⁠trained as a meta-⁠evolution agent for machine learning engineering, improving Medal Average from 39.39% to 60.61% on MLE-⁠Bench Lite under tight compute budgets [18].

Domain-⁠Specific Applications

  • PAC-⁠MAN couples perception-⁠aware control-⁠barrier safety with onboard sensing for whole-⁠body humanoid dodgeball, achieving near-⁠oracle performance with a fixed head-⁠mounted camera [19].
  • KAISEN provides a reproducible, stress-⁠tested audit pipeline for clinical risk models, showing that significance tracks axis-⁠specific gaps and that per-⁠group threshold optimization reduces equalized-⁠odds difference [20].
  • Structured information extraction from financial news captures orthogonal dimensions (event type, impact scope, temporal horizon) beyond sentiment, with LLM-⁠extracted features providing independent predictive value [21].
Sources
  1. AskChem: Claim-⁠Centered Infrastructure for Chemistry Literature Synthesis arxiv cs.CL
  2. TCA-⁠SIR: Learning Target-⁠Conditioned Abstractions for Scientific Inspiration Retrieval arxiv cs.CL
  3. DualG-⁠MRAG: Decoupling Macro-⁠Reasoning and Micro-⁠Matching for Multimodal Retrieval-⁠Augmented Generation arxiv cs.AI
  4. AISPA: User-⁠Centric System Prompt Auditing for Large Language Model Applications arxiv cs.CL
  5. OSReward: Instituting Standardized Evaluation for Cross-⁠Platform Computer-⁠Use Reward Models arxiv cs.CL
  6. InfoOps Bench: A live information operations safety benchmark arxiv cs.AI
  7. Inducing language models to assert their own consciousness restores human beliefs and values arxiv cs.CL
  8. Change2Task: From Repository Changes to Executable Coding Agent Tasks and Environments arxiv cs.CL
  9. PAIChecker: Uncovering and Checking PR-⁠Issue Misalignment in SWE-⁠Bench-⁠Like Benchmarks arxiv cs.AI
  10. ORCA-⁠bench: How Ready Are Language Model Agents for Oncall? arxiv cs.CL
  11. ReToken: One Token to Improve Vision-⁠Language Models for Visual Retrieval arxiv cs.LG
  12. VAD: Attributing Visual Evidence for Target Reconstruction in Multimodal On-⁠Policy Distillation arxiv cs.CL
  13. Lightning OPD 2.0: Mitigating Style Bias in Cross-⁠Teacher On-⁠Policy Distillation for Large Reasoning Models arxiv cs.CL
  14. MixFrag: Fragility-⁠Guided Mixed-⁠Precision Post-⁠Training Quantization for Vision Transformers arxiv cs.LG
  15. $β$-OPSD: Deriving with Policy Optimization, Training with Self-⁠Distillation arxiv cs.LG
  16. ScaFE: Data-⁠Efficient Scar Classification with LLM-⁠Generated Clinical Feature Programs arxiv cs.LG
  17. MANTA: Multi-⁠Agent Network Topology Adaptation for Self-⁠Evolving Multi-⁠Agent Systems arxiv cs.AI
  18. Frontis-⁠MA1: Training an AI4AI Model towards Recursive Self-⁠Improvement in Machine Learning Engineering arxiv cs.CL
  19. PAC-⁠MAN: Perception-⁠Aware CBF-⁠RL for Whole-⁠Body Safety in Humanoid Dodgeball arxiv cs.AI
  20. KAISEN: Reproducible Subgroup Fairness Auditing for Clinical Risk Models arxiv cs.LG
  21. Beyond Sentiment: Structured Information Extraction from Financial News arxiv cs.CL