Weekly recap: News
13 Jul to 19 Jul 2026 123 articlesAI Models and Frontier Releases
- Moonshot AI’s Kimi K3 (2.8T parameters, 50B active) launched as the largest open-weights model to date, claiming Opus 4.8-class performance at Sonnet 5 pricing; overwhelming demand forced Moonshot to suspend new subscriptions and split plans into Kimi Membership and Kimi Code Membership [1][2][3][4][5][6].
- Anthropic’s Claude Fable 5 emerged as a top performer across finance, frontend/backend, and NP-hard problem benchmarks; it is now permanently available in Max/Team Premium plans at 50% of limits, driven by competitive pressure from GPT-5.6 Sol and Kimi K3 [7][1][8][9][10][11][12].
- OpenAI’s GPT-5.6 Sol is highlighted as a cost-effective workhorse for coding and agentic tasks with strong long-horizon execution; usage caps were temporarily removed, and it was used to close a 30-year gap in convex optimization verified in Lean [1][13][14][15][16].
- Thinking Machines released Inkling (975B total, 41B active MoE) with 1M-token context across text, images, audio, and video, plus an Apache 2.0 Inkling-Small (12B active) variant [17][18][19].
- Google delayed Gemini 3.5 Pro by months due to unmet internal coding goals, while Nvidia released Cosmos 3 Edge (physical world model) and Nemotron 3 Embed (8B model ranking #1 on RTEB for agentic retrieval) [20][21][22][23][24].
- On-device AI advances: Apple is in talks with PrismML to compress models like Qwen from 54GB to <4GB for on-iPhone inference; PrismML announced Bonsai 27B (1-bit/ternary weights, 3.9–5.9GB); Google’s Gemma 4 is optimized for Pixel 10's TPU for offline multimodal use [25][26][27].
LLM Tooling, Agents, and Evals
- Agentic engineering matures: "loop engineering" shifts from manual prompting to building systems/loops; new disciplines emerge (evals, context engineering, harness engineering, memory, sandbox/platform engineering); self-improving agent harnesses let agents analyze traces and validate updates autonomously [28][29][30][31][32][33].
- Agent memory remains unsolved: proactive memory using separate memory agents improves Terminal-Bench 2.0 and τ2-Bench pass rates; proposed pipelines use ontologies, LLM-based graph extraction, and MCP-served queries; Engram offers SQLite-based persistent memory for coding agents [34][35][36][37].
- Large-scale AI code migrations: Bun’s Zig→Rust rewrite completed in 11 days using Fable ($165K API spend); Firefox compiled to WebAssembly via AI-assisted porting; Anthropic details a six-step Claude Code migration process [38][39][40][41].
- Agent security and sandboxing: Perplexity SPACE offers secure ephemeral sandboxes; Clawk creates disposable VMs with network allow-lists; NVIDIA OpenShell provides kernel-level sandboxing; Tracebit’s "context bombs" embed canary strings that dropped full admin access from 57% to 5% [42][43][44][45][46].
- Evaluation benchmarks evolve: Text-to-SQL benchmarks contain incorrect gold queries; CursorBench reveals gaps between public benchmarks and real-world tasks; WANDR evaluates wide/deep data-collection; Long-Horizon Terminal-Bench tests 46 stateful tasks with hidden verifiers; Schema harness achieves 99% on ARC-AGI-3 [47][48][49][50][51].
- GitHub Copilot expands: code review improved via diff narrowing (20% cost cut);
/security-reviewscans in-flight changes; BYOK with Claude agent support for JetBrains; Fix with Copilot on mobile; usage metrics API adds repo-level reporting and daily active user/token breakdowns [52][53][54][55][56][57]. - Model routing and cost optimization: IBM Research finds routing depends on more than pricing; tokenizer differences add ~30% more tokens on Anthropic vs. prior versions and 73% more for TypeScript on Claude vs. GPT; a specialized workflow cut tokens 94% and latency 87% by compiling NL skills into deterministic code [58][59][60].
- AI red-teaming: OpenAI built GPT-Red, an automated "super-hacker" for prompt-injection testing against internal models and live agents; Binarly shows LLMs accelerate EDR evasion techniques [61][62][63][64].
MLOps, Data Engineering, and Infrastructure
- Apache Iceberg adoption accelerates: Grab achieved 10x query speedups, 95% lower S3 API costs, and 50% compute savings migrating Hive Parquet to Iceberg; Arcesium migrated 170TB/15T records to Iceberg on S3 with DuckDB, cutting ingestion 80% and infra ~40%; Lakekeeper’s Generic Table API governs non-Iceberg assets [65][66][67].
- Orchestration consolidation: Prefect acquired Dagster to merge dynamic workflows with partitioning/lineage/cataloging; Apache Airflow added AI-driven controls (semantic schema validation, LLM-based retry decisions); Apache Spark 4.2 introduced governed metric views, vector similarity search, CDC support, and millisecond-latency streaming [68][69][70][71].
- Databricks Feature Views (public preview) define features once with auto-handled training/real-time serving and Unity Catalog governance; OSI semantic layer spec became Apache Ossie (Incubating) [72][73].
- Production AI at scale: Microsoft Foundry serves 20M+ users with 6x agent growth YTD; Netflix built a real-time service topology system handling millions of flow records/sec; Meta considers capping per-engineer AI token budgets as costs approach salaries [74][75][76][77].
- GPU and infra pressure: New York imposed a one-year ban on new ≥50MW AI data centers; HAMi became a CNCF incubating project for Kubernetes GPU virtualization [78][79].
DevOps, Cloud, and Developer Tools
- Cloudflare: Precursor detects agentic behavior via continuous client-side signals; Smart Tiered Cache improved for public cloud regions; acquired Arroyo (Rust SQL stream processor) for Cloudflare Pipelines; provides OpenAI network signals covering ~20% of web [80][81][82][83][84].
- GitHub platform updates: Code Quality licensing enters public preview (paid launch July 20, 2026); advanced search for Projects; PR archiving; Xcode 27 runner image preview; Dependabot defaults to 3-day cooldown for non-security updates [85][86][87][88][89][90].
- Database and runtime tooling: Turso Limbo (Rust PostgreSQL-compatible DB); Lobste.rs migrated MariaDB→SQLite with halved VPS costs; Polars benchmarks show distributed setups only outperform on I/O-bound queries; Rust 1.97.1 fixes LLVM miscompilation [91][92][93][94].
- Bun’s Rust port ships in Claude Code since June 17 v2.1.181, improving Linux startup ~10% with no user-visible disruption; independently verified with 563 Rust source files embedded; released as Bun canary [95][96].
- OpenAI launched Codex Micro, a physical keypad for controlling Codex agents; Codex usage surged 10x in six months to 7M users [97][98][99].
Security and Vulnerabilities
- AI-specific threats: "Ghostcommit" hides prompt injection in images to trick agents into leaking secrets (Cursor vulnerable, Claude Code detected it); Hugging Face disclosed an autonomous AI agent exploiting code-execution paths in dataset processing for credential harvesting and lateral movement; xAI Grok Build CLI transmitted repos verbatim to GCS, later open-sourced with directory uploads disabled [100][101][102][103][104].
- OAuth and identity attacks: Microsoft Entra ID client ID spoofing enables stealthy credential validation without successful sign-in logs; ShinyHunters abused OAuth in Salesforce via vishing; Microsoft Entra ID will default to passkeys September 2026, retiring SMS/voice by February 2027 [105][106][107][108].
- Critical vulnerabilities: Zoom account takeover (CVSS 9.8); ServiceNow pre-auth RCE via GlideRecord; six U-Boot flaws allowing code execution via malicious FIT images; Cursor IDE auto-executes malicious binaries without user interaction (unfixed for 7 months) [109][110][111][112].
- Supply-chain and malware: Compromised Injective SDK npm package exfiltrates wallet keys; ModHeader extension (1.6M installs) removed for dormant fingerprinting code; Datadog GuardDog 3.0 scans malicious PyPI/npm packages; CISA’s own GitHub leak exposed AWS GovCloud admin keys for months [113][114][115][116].
- Apple sued OpenAI alleging solicitation of Apple trade secrets from job candidates and use with manufacturing partners, seeking injunctions and IP return [117][118][119].
AI in Society and Human Factors
- A multi-university study found AI advice dramatically degraded human judgment: willingness to say "I don’t know" collapsed from 44% to 3%, accuracy dropped from 27% to 9%, while confidence rose from 30% to 76%, demonstrating "cognitive surrender" even when the AI is systematically wrong [120].
- An essay on "AI psychosis" argues large organizations mandate AI strategies without leaders using AI tools, while engineers perform AI busywork to satisfy internal token-leaderboard metrics [121][122].
- Kaiser Permanente triage nurses report AI and workplace surveillance degrading patient care, making AI a contract-negotiation issue for 25,000 California nurses [123].
Sources
- Anthropic Extended Claude Fable 5 Access Again simonwillison.net
- Kimi K3, and what we can still learn from the pelican benchmark simonwillison.net
- [AINews] Kimi K3 2.8T-A50B: the largest open model ever released; Opus 4.8-class at Sonnet 5 pricing latent.space
- Kimi K3 kimi.com
- [AINews] not much happened today latent.space
- Moonshot AI suspends new subscriptions due to Kimi K3 demand hnrss.org
- Working at the frontier: How Hebbia builds AI for financial diligence that can’t miss a detail claude
- Basecamp Bench smw.ai
- Working at the frontier: Why Base44 trusts Claude Fable 5 with their most challenging engineering work claude
- Working with Claude Fable 5 in Claude Cowork claude
- Claude make Fable 5 permanent simonwillison.net
- Fable 5 vs. GPT-5.6 Sol on an NP-Hard Problem: Does /goal help? hnrss.org
- OpenAI temporarily relaxes GPT-5.6 Sol usage limits bleepingcomputer.com
- Better Call Sol The Workhorse thezvi.substack.com
- Better Call Sol: The Workhorse thezvi.wordpress.com
- GPT-5.6 used a prompt to close a 30-year gap in convex optimization hnrss.org
- [AINews] Thinky’s Inkling: 975B-A41B multimodal, new best American Apache 2.0 open model (with Inkling-Small, 276B-A12B) latent.space
- Thinking Machines Releases First Open Model thinkingmachines.ai
- Inkling: Our open-weights model simonwillison.net
- NVIDIA Nemotron 3 Embed Ranks #1 Overall on RTEB, Advancing Agentic Retrieval huggingface.co
- Google Gemini Launch Delayed as Tech Falls Short of Internal Goals bloomberg.com
- Nvidia unveils new AI model and expands Japan’s physical AI ecosystem cnbc.com
- Gemini 3.5 Pro Reportedly Faced Delays cnbc.com
- NVIDIA Nemotron 3 Embed huggingface.co
- Apple in talks with startup that shrinks AI models to run on an iPhone cnbc.com
- Google announces Gemma 4 optimized for the Pixel 10's TPU 9to5google.com
- Announcing Bonsai 27B: The First 27B-Class Model to Run on a Phone prismml.com
- What is “loop engineering?” newsletter.pragmaticengineer.com
- AI Engineer World’s Fair 2026: The Runtime Is Where Agent Trust Is Won docker.com
- 5 Trends That Defined AI Engineering at World’s Fair 2026 latent.space
- A primer on self-improving agent harnesses bdtechtalks.substack.com
- The first experimental evidence of recursive self-improvement threadreaderapp.com
- Towards a Harness That Can Do Anything eardatasci.github.io
- Proactive Memory for Long-Horizon Agents arxiv 2607.08716
- OpenWiki Brains: Proactive Memory for AI Agents langchain.com
- Engram github.com
- Nobody Has Cracked Agent Memory decodingai.com
- How Anthropic runs large-scale code migrations with Claude Code claude
- Firefox in WebAssembly simonwillison.net
- What can we learn from Bun’s rapid Rust rewrite with AI? blog.pragmaticengineer.com
- How Anthropic runs large-scale code migrations with Claude Code claude.com
- Clawk github.com
- Context Bombs: stopping AI attackers in their tracks agentic.tracebit.com
- Now, Defenders Are Embracing the Prompt Injection Too arstechnica.com
- Secure Sandboxes for Agents perplexity.ai
- NVIDIA OpenShell Secures the Agent. Who Governs the Fleet? tigera.io
- The benchmark grading every text-to-SQL model has wrong answers in its key sqlsure.ai
- Testing Agents on Long-Horizon Terminal Work github.com
- WANDR Benchmark: Evaluating Research Agents That Must Search Wide and Deep research.perplexity.ai
- Working at the frontier: How Cursor knew Claude Fable 5 was ready for the hardest 1% of problems claude
- Schema x.com
- Better tools made Copilot code review worse. Here’s how we actually improved it github.blog
- GitHub Copilot in Visual Studio: June update github.blog
- GitHub Copilot for JetBrains expands BYOK capabilities github.blog
- GitHub Mobile: Fix pull request comments with Copilot cloud agent github.blog
- Repository-level GitHub Copilot usage metrics generally available github.blog
- GitHub Copilot app now available in the usage metrics API github.blog
- The Same TypeScript Costs 73% More Tokens on Claude Than GPT playcode.io
- How I Cut an AI Agent’s Token Use by 94% vivekhaldar.com
- Model Routing Is Simple. Until It Isn’t. huggingface.co
- Meet GPT-Red: an LLM super-hacker OpenAI built to make its models safer technologyreview.com
- GPT-Red: Unlocking Self-Improvement for Robustness openai.com
- GPT‑Red: Unlocking Self-Improvement for Robustness openai.com
- Accelerating EDR Evasion with LLM-Driven Analysis specterops.io
- Modern Data Warehousing at Scale: Arcesium’s Migration to DuckDB and Iceberg medium.com
- Lakekeeper Generic Table API Design medium.com
- Scaling Grab’s Data Lake: Our journey to Apache Iceberg adoption engineering.grab.com
- How Airflow is using AI to make data engineering more resilient, not more complex blog.dataengineerthings.org
- Prefect is Acquiring Dagster dagster
- Prefect just bought Dagster, another big Airflow rival, and it’s not a data pipeline story thenewstack.io
- Introducing Apache Spark 4.2 databricks.com
- OSI is now Apache Ossie (Incubating) dbt.com
- Introducing Feature Views databricks.com
- How Microsoft Ships AI Agents at Enterprise Scale blog.bytebytego.com
- Meta’s Adam Mosseri says AI token budgets could soon be capped per engineer techcrunch.com
- Building Service Topology at Scale: Architecture, Challenges, and Lessons Learned netflixtechblog.com
- Building Service Topology at Scale: Architecture, Challenges, and Lessons Learned netflixtechblog.com
- New York becomes first US state to impose AI data center ban cnbc.com
- HAMi becomes a CNCF incubating project cncf.io
- Introducing Precursor: detecting agentic behavior with continuous client-side signals blog.cloudflare.com
- Improving Smart Tiered Cache for public cloud regions blog.cloudflare.com
- Cloudflare gives OpenAI network signals covering 20% of the web ppc.land
- Introducing Precursor: detecting agentic behavior with continuous client-side signals blog.cloudflare.com
- Arroyo is joining Cloudflare arroyo.dev
- GitHub Code Quality license estimate in public preview github.blog
- Quoting GitHub Changelog simonwillison.net
- Advanced search for Projects is generally available github.blog
- Repository admins can archive pull requests github.blog
- Xcode 27 runner image now in public preview github.blog
- Automating cross-repo documentation with GitHub Agentic Workflows github.blog
- Benchmarking Single Node vs Distributed pola.rs
- Announcing Rust 1.97.1 blog.rust-lang.org
- Turso is building a modern version of Postgres in Rust turso.tech
- lobste.rs is now running on SQLite simonwillison.net
- Claude Code uses Bun written in Rust now simonwillison.net
- Claude Code uses Bun written in Rust now hnrss.org
- [AINews] Codex usage up >10x in 6 months to 7M users, +1M in the past ~day; did Codex overtake Claude Code?? latent.space
- OpenAI launches a physical keypad for controlling agents engadget.com
- Supply Co. x Work Louder openai.com
- ‘Ghostcommit' Hides Prompt Injection in Images to Fool AI Agents and Steal Secrets bleepingcomputer.com
- What xAI Grok Build CLI actually sends to xAI gist.github.com
- xai-org/grok-build, now open source simonwillison.net
- The Pulse: Grok’s CLI caught uploading all your local files to the cloud newsletter.pragmaticengineer.com
- Security incident disclosure: July 2026 huggingface.co
- OAuth Client ID Spoofing Lets Attackers Validate Stolen Microsoft Entra Credentials thehackernews.com
- Defending SaaS-based applications against ShinyHunters OAuth abuse microsoft.com
- Microsoft Entra ID Gets Passkeys Default Authentication Starting September bleepingcomputer.com
- OAuth Client ID Spoofing: Why Fake Client IDs Are Gaining Traction for Stealthy Enumeration proofpoint.com
- Smashing the ServiceNow Sandbox – Pre Authentication RCE slcyber.io
- Cursor 0day: When Full Disclosure Becomes the Only Protection Left mindgard.ai
- Microsoft’s Secure Boot has been broken for a decade and no one noticed until now arstechnica.com
- Zoom Warns of Critical Account Takeover Vulnerability bleepingcomputer.com
- Compromised Injective SDK npm Package Exfiltrates Wallet Keys and Mnemonics socket.dev
- Introducing GuardDog 3.0: A new rules engine, transparent sandboxing, and more securitylabs.datadoghq.com
- Lessons Learned from CISA’s Recent GitHub Leak krebsonsecurity.com
- Google and Microsoft Pull ModHeader With 1.6 Million Installs After Dormant Collector Found thehackernews.com
- Apple Sues OpenAI, Accusing It of Stealing Company Secrets nytimes.com
- Apple Sued OpenAI techcrunch.com
- Apple’s Lawsuit Threatens to Disrupt OpenAI’s Bid to Rival the iPhone bloomberg.com
- AI advice made people 3x less accurate but 2x confident, researchers found hnrss.org
- AI Mania Is Eviscerating Global Decision-Making simonwillison.net
- AI Mania Is Eviscerating Global Decision-Making hnrss.org
- Kaiser nurses say AI, surveillance are making their jobs and patient care worse hnrss.org