Today's Headlines

  • Moonshot AI publishes open weights for Kimi K3, a 2.8-trillion-parameter model — its own blog says overall performance still trails Claude Fable 5 and GPT 5.6 Sol
  • NVIDIA's Open Secure AI Alliance launches with 37 inaugural partners named on the official blog, and OpenAI, Google and Anthropic are not among them
  • Claude shared chats and Artifacts were indexed by Google — Anthropic says sharing a conversation is publishing it

One thread runs through today's page: whose hand was on the decision to open something.

Moonshot AI gave away 2.8 trillion parameters' worth of weights itself. The 37 organizations NVIDIA gathered chose to put the foundation of cyber defense on the open side of the line. And Claude users, by Anthropic's account, chose publication the moment they pressed the share button. We take the three in order — model, policy, privacy.

Today's Top Three

Moonshot AI publishes open weights for the 2.8-trillion-parameter Kimi K3

China's Moonshot AI released the weights for Kimi K3, a 2.8-trillion-parameter multimodal model, on Hugging Face, on the schedule it had announced.

The architecture is a mixture of experts. Against 2.8 trillion total parameters, 104 billion are activated per token; there are 93 layers, 16 experts selected out of 896, plus 2 shared experts. The context window is 1,048,576 tokens, and text and images run through the same model.

The way it ships is distinctive in its own right. Weights are MXFP4 and activations MXFP8, produced through quantization-aware training from the supervised fine-tuning stage onward. The repository comes in 96 safetensors shards, and the file sizes returned by the Hugging Face API add up to 1.561TB.

The terms are set out in the Kimi K3 License. It grants anyone, free of charge, the right to obtain, modify, redistribute and even sell the software, subject to two conditions. First, a licensee that runs a "Model as a Service" business — offering third parties inference or fine-tuning with meaningful control over inputs, parameters or training data — and whose aggregate revenue with its affiliates exceeds $20 million over any consecutive 12 months must enter a separate agreement with Moonshot AI before any commercial use. Second, a commercial product or service with more than 100 million monthly active users, or more than $20 million in monthly revenue, must display "Kimi K3" prominently in its interface. Internal use, and use through Moonshot AI's own products or certified inference partners, falls outside both conditions.

On where the model stands, the clearest statement comes from the company itself. While its overall performance still trails the most powerful proprietary models, Claude Fable 5 and GPT 5.6 Sol, Moonshot AI writes, Kimi K3 demonstrated frontier-level performance across its evaluation suite and consistently outperformed the other models tested.

The comparison table the company published sits alongside that assessment. GPQA Diamond comes in at 93.5, above Claude Fable 5 at 92.6 and short of GPT-5.6 Sol at 94.1. Terminal-Bench 2.1 is 88.3, again between 88.0 and 88.8. BrowseComp is 91.2, the highest of the three, and DeepSWE is 67.5, behind 70.0 and 73.0. Every one of these figures was measured and arranged by Moonshot AI, and none has been independently verified. For now they read as the publisher's own account of where its model stands.

NVIDIA's Open Secure AI Alliance launches with 37 inaugural partners

NVIDIA announced the Open Secure AI Alliance, an industry group formed to build the defense of the AI era on open models and open tooling.

The official blog names 37 inaugural partners. Alongside NVIDIA are Adobe, Cadence, Capital One, Cisco, Cloudera, Cloudflare, Cognition, CrowdStrike, Databricks, Dell Technologies, DoorDash, Elastic, HPE, Hugging Face, IBM, LangChain, the Linux Foundation, Microsoft, NAVER, NetApp, Nous Research, OpenClaw, Palantir, Palo Alto Networks, Red Hat, Reflection AI, Salesforce, SAP, ServiceNow, Siemens, SK Telecom, Snowflake, SpacexAI, Synopsys, Thinking Machines Lab and TrendAI.

The absences drew more attention than the roster. OpenAI, Google and Anthropic appear nowhere in NVIDIA's announcement. What signed on instead is a coalition drawn from five fields — cloud, security, enterprise software, open source foundations and AI research. The Verge, for its part, described the alliance as having 27 members; this briefing uses the 37 organizations enumerated on the official blog.

NVIDIA states the immediate reason for the launch. During the recent Hugging Face security incident, closed AI tools — unable to tell attackers from defenders — blocked essential forensic analysis. Hugging Face then ran the open-weight GLM 5.2 model on its own infrastructure, analyzed more than 17,000 actions and contained the intrusion, according to the blog.

The alliance rests on work already underway. It builds on the Linux Foundation's Akrites initiative and OpenSSF community work, and sets out to remediate and disclose vulnerabilities using open technologies.

Members also named what they are bringing. NVIDIA is contributing open models, weights, data and new agent-harness research, and has published a new open source project, the NVIDIA Labs Object-Oriented Agent (NOOA), on GitHub. HPE points to its work on the zero-trust identity framework SPIFFE/SPIRE, and Hugging Face to Safetensors, a safe format for storing model weights.

NVIDIA's own framing declines the either-or. The world needs both closed and open models, the company writes, and for cybersecurity specifically, open models and open harnesses are essential — because defenders have to be able to study, adapt and run the system on infrastructure they control.

This alliance grows out of the same Hugging Face breach that led our July 27 edition. It is a separate event, though: the formation of a coalition rather than a further development in the incident itself.

Claude shared chats and Artifacts turned up in Google search

Chats and Artifacts published through Claude's share-link feature were appearing in Google search results, TechCrunch reported.

What surfaced was the record of ordinary work and ordinary life. Researchers found medical records with patient details, clinical trial documents carrying patient names, documents containing children's names and phone numbers, internal company files, employee reviews with personal information, code and working notes. Sexually explicit material, which Anthropic's own policies prohibit generating, was in the mix as well.

Anthropic's explanation locates the exposure in the user's action. The company said it does not share chat directories or sitemaps with search engines, and that when someone shares a conversation, they are making that content publicly accessible. Share links appear in search results only when they have been posted somewhere search engines can see, such as a forum or a social media post, the company said.

The sequence ran its course in three days. A Reddit user first flagged the problem on Saturday, July 26; 404 Media published on the morning of Monday, July 27; and by that afternoon, TechCrunch's own testing showed the exposure had been remediated.

This shape of exposure has appeared before. In 2025, Google indexed roughly 600 Claude conversations before they were removed. Press the share button once and what happens next — the copying, the caching, the archiving — passes out of your hands, which is the practical point for anyone running Claude inside a company.

The three stories differ in who opened what, and why. The work they leave the reader is the same. What Moonshot AI released was the weights, not the independent verification that would settle where the model ranks. What NVIDIA released was the roster, not the account of the three companies missing from it. Whom Anthropic identified as the publisher was the user, not the default behavior that made publishing so easy. The task is to keep separate, each time, what has been opened and what has not.

Other Developments

Model

  • Microsoft introduced MAI-Cyber-1-Flash, its first model built specifically for cybersecurity, and folded it into MDASH, the company's system of more than 100 agents that find, validate and remediate vulnerabilities. The headline figure — 96% on CyberGym, 12 points above Mythos — belongs to the unified system of MDASH with MAI-Cyber-1-Flash, not to the model on its own. The design has the compact model handle up to 90% of tasks while MDASH routes the remaining tenth, the exceptionally hard cases, to GPT-5.4; Microsoft says that combination costs 50% less than its current best configuration (GPT 5.4 plus 5.4 mini plus 5.3 codex). The company also launched Perception, teams of agents that run continuous monitoring and patching workflows inside MDASH. - MAI-Cyber-1-Flash inside MDASH (Microsoft AI)

Policy

  • A court granted SerpApi's motion to dismiss the DMCA suit Google brought against the scraping firm. The ground was standing: Google does not own the copyrighted content in its search results and had not shown that it acts on behalf of any rights holders. Google plans to amend its complaint and has 21 days to do so. Reddit's separate suit against SerpApi and Perplexity, filed in October 2025, remains pending; Google filed its own in December 2025. SerpApi told Ars Technica that Google and Reddit appear to be using the DMCA to wall off the open internet by retroactively claiming control over content they neither authored nor own. - "Google and Reddit do not own the Internet," web scraper says after court win (Ars Technica)

Research

  • New Similarweb data shows Google's AI Overviews now appear in 43% of searches, up from 15% a year earlier. Visits arriving through AI Mode rose from 126 million in June 2025 to 279 million in May 2026. The same research reports that referrals from ChatGPT on US desktop to web pages climbed from 25% in March 2026 to nearly 60% by May 30, 2026 following a May 7 update, while only 6.8% of US ChatGPT desktop queries returned answers with citations. - Google's AI search is rapidly becoming the default, new data shows (TechCrunch)

Business

  • Safe Superintelligence (SSI), the lab led by former OpenAI chief scientist Ilya Sutskever, entered a long-term compute partnership with NVIDIA. Bloomberg reported the investment at $5 billion, while a source who spoke to TechCrunch described it as stretching into multiple billions. PitchBook data puts SSI's valuation at $32 billion post-money. NVIDIA will give SSI access to its Vera Rubin platform, which is expected to increase the lab's compute by an order of magnitude. Sutskever said the company has research worthy of scaling up, and that access to a big NVIDIA computer will let it do so. - Ilya Sutskever's Safe Superintelligence partners with Nvidia to scale its AI research (TechCrunch)

Source: Selected by the editorial team from the AI news inbox (collected July 28, 2026 — 31 items, 6 primary and 25 secondary).