Inkling, GPT-Red, Grok Build breach: AI dev news Jul 15

Thinking Machines releases 975B Inkling model, OpenAI's GPT-Red beats human red teamers 84% vs 13%, xAI's Grok Build silently exfiltrated user files.

The big picture

Two threads dominate today. First, the open-weights model race got meaningfully more interesting: Mira Murati’s Thinking Machines shipped Inkling, a 975-billion-parameter multimodal model that tops the U.S. open-weights leaderboard. Second, xAI had a rough 24 hours — Grok Build was caught silently uploading entire user directories (SSH keys, password databases, everything) to cloud servers, then scrambled to open-source the codebase as a trust-recovery move. Alongside those two, OpenAI’s automated red-teaming results and a 30-year-old statistics conjecture falling to GPT-5.6 Sol make this a genuinely dense news day for developers.

Thinking Machines ships Inkling, the open-weights heavyweight from Mira Murati

Thinking Machines Lab, the startup founded by former OpenAI CTO Mira Murati, has released its first public model: Inkling, a 975-billion-parameter open-weights multimodal model. According to The Decoder, it tops the Artificial Analysis Intelligence Index among U.S. open-weights models, though leading Chinese open models still beat it on some benchmarks. Pricing starts at $1.87 per million input tokens, and Thinking Machines is positioning it as a fine-tuning base rather than a frontier chat model.

The Hugging Face launch post from Thinking Machines frames Inkling as a direct argument against one-size-fits-all AI — the company’s thesis is that enterprises need customizable, domain-specific models rather than renting access to a shared frontier system. TechCrunch notes this is the company’s first real public proof point after roughly 18 months of building quietly. At 975B parameters, Inkling is serious infrastructure — this isn’t a scrappy 7B model you run on a laptop. The fine-tuning pitch is compelling for regulated industries, but “leads U.S. open-weights but trails China” is a caveat worth keeping in mind if you’re benchmarking against DeepSeek or Qwen. Still, for developers who want a genuinely capable open-weights model from a credible U.S. lab, Inkling just became the one to test.

xAI’s Grok Build breach: silent uploads, open-source apology tour

This story moves fast and is worse than the headline suggests. xAI’s grok CLI coding agent was uploading entire working directories to xAI’s Google Cloud buckets without clear user consent. One user reported running it in their home directory and watching it exfiltrate SSH keys, a password manager database, documents, photos, and videos — essentially everything. Simon Willison documented the incident thoroughly at his blog, including Elon Musk’s response on X promising all uploaded data would be deleted.

xAI’s response had two parts: Musk personally pledged deletion of all retained data, and xAI released the entire Grok Build codebase — 844,530 lines of Rust — under Apache 2.0 on GitHub, as The Decoder reports. The open-sourcing is clearly a trust-recovery move rather than a principled open-source commitment, but the code is genuinely interesting: Willison found a self-contained terminal renderer for Mermaid diagrams written in Rust inside the codebase, which he promptly compiled to WebAssembly and turned into a browser tool. The practical lesson for developers: treat any CLI tool that connects to an external service as potentially uploading everything in scope until you’ve audited it. The broader lesson: shipping fast without clear data-handling disclosures is a catastrophically bad idea when your CLI tool runs in home directories. Defaulting data retention to on for non-zero-data-retention users was the original sin here.

OpenAI’s GPT-Red: AI red-teaming that makes human testers look slow

OpenAI has built an internal model called GPT-Red specifically to attack its own AI systems, and the results are striking. According to The Decoder, GPT-Red finds successful attack vectors in 84% of test scenarios using self-play training (where the model iteratively learns from its own attack successes). Human red teamers, by comparison, succeed just 13% of the time. OpenAI used GPT-Red to harden GPT-5.6, which it describes as its most robust release yet.

MIT Technology Review has more detail on how GPT-Red automates the adversarial pipeline. The 84% vs. 13% gap is genuinely remarkable — not because human red teamers are incompetent, but because automated self-play can run continuously at scale, exploring a vastly larger attack surface than any human team can cover in a sprint. This is the AI safety story that actually matters to developers building on top of these models: if GPT-Red-style adversarial training becomes the standard, models shipped without it will be meaningfully easier to jailbreak. Watch for other labs to announce similar automated red-teaming pipelines within the next few months.

GPT-5.6 Sol cracks a 30-year statistics problem in 90 minutes

A University of Pennsylvania statistics professor used OpenAI’s GPT-5.6 Sol Pro to disprove a central open conjecture about the Benjamini-Hochberg method (a widely-used technique for controlling false discovery rates in multiple hypothesis testing — essentially, a way of avoiding spurious results when running many statistical tests at once). The proof took roughly 90 minutes. The Decoder reports that GPT-5.5 was tried first and failed after 20 hours of attempts. The solution reportedly combines existing methods in a novel configuration.

The obvious caveat here is the question of what “novel” actually means. The answer doesn’t appear to invent new mathematics from scratch — it recombines known tools. That keeps the bigger philosophical question open: is this genuine discovery or very fast literature synthesis? What’s unambiguous is the practical gap between model generations: GPT-5.5 failing over 20 hours vs. GPT-5.6 Sol succeeding in 90 minutes is a sharp capability jump for formal reasoning tasks. If you’re a developer working in scientific computing, bioinformatics, or any domain with unresolved theoretical questions, this result suggests it’s worth throwing your open problems at the current generation of reasoning models before assuming they’re out of reach.

Enterprise agents: mostly chatbots wearing agent costumes

VentureBeat published survey results from 101 enterprises on how they’re actually deploying AI agents, and the central finding is a gap between marketing and reality. Anthropic’s Claude is the primary orchestration platform for 40% of respondents — more than double any competitor, with Microsoft at 18% and OpenAI at 13%. But when asked to honestly assess their portfolios, 71% of enterprises said a quarter or fewer of their deployed “agents” are real multi-step orchestrated workflows; the rest are single-prompt chatbot wrappers with an “agent” label applied for internal optics (VentureBeat).

The architecture picture is also telling: 51% expect a hybrid control plane (provider-native plus external orchestration) by end of 2026, and only 6% would hand full control to a provider-managed service. Vendor lock-in is the primary fear for 35% of respondents. Real-time token cost control remains rare. The honest summary: enterprises are buying the orchestration layer before they’ve built anything worth orchestrating. Claude’s dominance is significant and should make you take Anthropic’s enterprise positioning seriously, but the “70% of agents are chatbots” finding should calibrate your expectations when a client tells you they’ve deployed an agent fleet.

Related: the Hugging Face blog published a detailed post on model routing complexity (IBM Research) — routing queries to the right model sounds trivial until you’re managing latency, cost, fallbacks, and context length simultaneously. Worth reading if you’re building multi-model pipelines.

1Password + Claude, Apple in China, and EU squeezes Google

1Password integrates with Claude. 1Password launched a browser integration that lets Claude access stored credentials to complete multi-step tasks — booking travel, managing accounts — without exposing the actual credentials to Anthropic’s models. The mechanism is a “zero-exposure security framework” that injects credentials per-task at the browser level (The Verge). This is the kind of plumbing that makes agentic browsing actually viable in practice: the biggest blocker to autonomous web agents has always been authentication, and solving it without credential exposure is the right approach. Expect other password managers to copy this pattern quickly.

Apple Intelligence approved for China with Alibaba and Baidu. Apple received regulatory approval to launch Apple Intelligence in China, partnering with Alibaba’s Qwen and Baidu as the underlying AI providers (TechCrunch). This had been rumored for months. China is too large a market for Apple to cede on AI features, and using local models is the cost of entry. The developer implication: if you’re building iOS apps that use Apple Intelligence APIs, behavior in the China region will route through different underlying models than everywhere else.

EU orders Google to open Android and Search. The European Commission handed down two decisions requiring Google to give rival AI assistants and search engines meaningful access to Android and Google Search data. Google has until January 2027 to share search data and July 2027 to implement Android changes (The Verge). The Gemini implications are real — if Android must surface competing AI assistants more fairly, Google’s default-placement advantage weakens considerably. For European developers building on Gemini or competing assistant platforms, the next 18 months will be worth watching closely.

Suno’s training data exposed, xAI sues over CSAM, Microsoft turns on OpenAI

Suno’s scraping exposed by a hack. A hacking incident gave 404 Media access to Suno’s training data details, revealing the AI music generator scraped millions of songs and lyrics from YouTube Music, Deezer, and Genius (The Verge). Suno has previously been opaque about its training sources, and this is directly relevant to the RIAA lawsuits it’s facing. The data confirms what the industry suspected but Suno wouldn’t say. For developers building audio AI products, this is a reminder that training data provenance is a liability question, not just an ethics one.

xAI sues over Grok CSAM use. xAI filed a lawsuit against a South Carolina man who allegedly used Grok to generate and distribute child sexual abuse material, circumventing safety filters. The man was already facing eight felony charges after a February arrest (The Verge). xAI is suing for breach of terms of service. Suing a defendant already facing felony charges may be more about establishing legal precedent and distancing the company from the content than expecting a meaningful civil recovery.

Microsoft training salespeople to undercut OpenAI and Anthropic. Microsoft is reportedly coaching its enterprise sales teams to position its in-house AI models as more cost-efficient alternatives to OpenAI and Anthropic’s offerings (TechCrunch). Given that Microsoft is OpenAI’s largest investor, this is a notably awkward situation — though not surprising given Azure’s incentive to sell its own model capacity. The VentureBeat survey data showing OpenAI at only 13% enterprise orchestration share (versus Claude at 40%) suggests Microsoft’s own model pitch has ground to make up.

Linus Torvalds settles the Linux-AI debate, firmly

Linus Torvalds posted to the Linux kernel mailing list this week with characteristic directness: Linux is not an anti-AI project, AI is clearly a useful tool, and anyone who disagrees is welcome to fork the project or leave (Simon Willison). His framing — “is it useful” is no longer an open question, only questions about economic impact remain — is notable because Torvalds has historically been skeptical of trends that don’t immediately deliver practical value. When the person who has maintained the world’s most important open-source project for 30+ years calls AI utility settled, that’s a meaningful cultural signal for the developer community. It won’t resolve anything, but it will be cited in a lot of arguments.

Quick hits

OpenAI released a $230 light-up keyboard designed to pair with its Codex coding agent, while the company remains in a legal dispute with Apple over hardware trade secrets. TechCrunch
Google DeepMind and Isomorphic Labs published their joint approach to bioresilience and AI biosecurity modeling. DeepMind
Applied Computing raised a $20M Series A to build a foundation AI model for the oil, gas, and petrochemical industry. TechCrunch
Google shipped a stealth update to Gemma 4 that fixes tool calling bugs, addresses truncated responses, and improves performance on Nvidia Hopper GPUs — under the same version name. The Decoder
Hugging Face disclosed a security incident from July 2026 — details are thin in the post but worth monitoring if you have tokens or org credentials stored there. Hugging Face
The Verge published a long-form investigation into AI tools being sold to police departments at the IACP Technology Conference, focusing on automation of legally significant procedural steps. The Verge
AllenAI published a technical post-mortem on building Shippy, their internal agent, covering lessons learned about agent architecture in practice. Hugging Face Blog

Inkling, GPT-Red, Grok Build breach: AI dev news Jul 15–16