Top Stories
Claude shared links showed up in Google search
A 404 Media report says Claude users’ public share links were appearing in Google search results, making conversations and creations readable by strangers. Its headline and dek say users may not have realized that creating a public link could put their chats in search results. 404 Media reported the discovery.
TechCrunch found that a search for site:claude.ai/share returned shared conversations and Artifacts. The examples it reviewed included health records, private company documents, children’s names and phone numbers, code, and work notes. Claude’s sharing interface warns that “Anyone with the link can view” a conversation or project. TechCrunch’s report describes the available private and public-link choices. The episode lands in a privacy story the site has been tracking for months, including our chatbot data-collection audit and earlier reports on Google indexing shared AI content.
Anthropic said it does not provide chat directories or sitemaps to search engines and that share links appear in results only after users post them somewhere crawlers can see. Google said site owners control crawl and index directives. TechCrunch said its test search returned no results by Monday afternoon and directed users to Settings, Privacy, and Shared Chats to review existing links. The same report carries both responses and the audit steps.
Moonshot posts full Kimi K3 weights with a conditional license
Moonshot’s Hugging Face model card lists Kimi K3 as an open-weight release with 2.8 trillion total parameters, 104 billion activated parameters, and a 1,048,576-token context length. It describes a multimodal mixture-of-experts model built around Kimi Delta Attention and Attention Residuals, with 16 of 896 experts activated. The model card lists the architecture and release details.
The card lists Transformers and Safetensors, along with vLLM, SGLang, and TokenSpeed as supported deployment tools. It does not give a minimum GPU or VRAM requirement, so the weight release alone does not answer whether a particular local machine can run it. Moonshot’s model page is the source for those deployment details.
The Kimi K3 license allows copying, modifying, distributing, and selling the software, but adds commercial conditions. A Model-as-a-Service business whose licensee and affiliates exceed $20 million in aggregate revenue over any consecutive 12 months must make a separate agreement with Moonshot before commercial use. Products above 100 million monthly active users or $20 million in monthly revenue must prominently display the Kimi K3 name. The license sets out both thresholds. For context on how this fits Moonshot’s earlier open-weight pace, see our coverage of Kimi K2.6 storming the rankings.
Microsoft pairs a cyber model with a larger fallback
Microsoft says MAI-Cyber-1-Flash is a compact, code-heavy security model derived from the MAI-Thinking-1 lineage and integrated into MDASH, its multi-agent vulnerability identification and remediation harness. The company says the model is designed to handle up to 90% of MDASH tasks, leaving the 10% it calls exceptionally hard for GPT-5.4. Microsoft AI describes the model and routing.
Microsoft reports that the combined MDASH system scored 95.95% on CyberGym and that the configuration cuts costs by 50% against its previous MDASH offering. Those are Microsoft’s own benchmark and cost claims, not independent results. The model announcement gives the score and comparison.
Microsoft also announced Project Perception, which coordinates red-team agents that look for compromise paths, blue-team agents that assess risk, and green-team agents that take corrective action. The company says public preview starts August 3. Microsoft’s security blog describes the system and date.
Anthropic opposes a blanket open-weight ban
Anthropic CEO Dario Amodei writes that the company has never advocated for a ban on open-weight models. He supports tighter controls on advanced chips and chipmaking equipment reaching China, along with action against industrial-scale distillation operations. Anthropic’s position statement lays out the alternatives.
Amodei instead calls for mandatory pre-release testing for sufficiently capable models, whether their weights are open or closed. The testing he names covers cyber, biological, and alignment risks, while less capable models from startups and academia would be exempt. The statement says the testing should apply regardless of model origin or openness.
Nvidia and Safe Superintelligence announce a compute partnership
Nvidia says it has invested in Safe Superintelligence and will give the lab access to its next-generation Vera Rubin platform. The release says that access will let SSI increase its compute by an order of magnitude, while the two companies collaborate on current and future Nvidia platforms. Nvidia announced the partnership.
The announcement does not disclose the investment amount or describe a model. It describes SSI as a lab founded by Ilya Sutskever and Daniel Levy in 2024, with a single goal of building what it calls safe superintelligence. Those research and scaling statements come from Nvidia and SSI, not from an independent evaluation. The release provides the partnership’s public details.
Nvidia releases a fast simulator for surgical robotics
Nvidia describes Cosmos-H-Dreams as an action-conditioned generative simulator for surgical robotics, specialized in the released version for da Vinci Research Kit tabletop suturing. The post reports interactive operation at about 160 frames per second on one Nvidia RTX PRO 6000 GPU. The Hugging Face post explains the simulator and its reported speed. The release joins earlier coverage of autonomous surgical-robot research.
The release includes links to code, a checkpoint, a dataset, and a fine-tuning recipe. Nvidia says Cosmos-H-Dreams is a research and development platform, not a diagnostic system, a replacement for intraoperative imaging, or a controller for a physical surgical robot. The same post lists the releases and limitations.
Quick Hits
- Spotify’s AI-music problem: SoullessMusic and SlopTracker are tracking suspected AI artists because Spotify did not show AI labels on the dozens of tracks reviewed by 404 Media. The report also warns that detection can produce false positives. 404 Media has the details.
- Meta AI enters Threads DMs: Meta is rolling out private Threads conversations globally, including the ability to send posts, images, links, and videos for follow-up questions. Users can mute Meta AI, mark posts “Not interested,” or hide its replies. TechCrunch reports on the rollout.
- Data-center opposition meets aggressive policing: Emporia, Kansas, police arrested high-school physics teacher Lux Claridge after a clap at a city meeting; commissioners voted 5-0 on zoning for the proposed 1,000-acre Flint Hills Digital Campus. 404 Media reports.
- Enigma emerges with $71 million: The startup says it has more than 100 robots and is opening an online experiment so people around the world can try controlling them. TechCrunch covers the seed round and experiment.
Worth Watching
- Claude’s sharing cleanup: TechCrunch’s follow-up search returned no results, but users still need to audit old public shares. The report lists the review path.
- Kimi K3 outside the hosted API: The model card makes the weights and deployment tooling public, while the license places conditions on larger Model-as-a-Service businesses. Independent tests of real hardware requirements and performance will matter next. The model card and license define the starting point.
- Labels versus detection on Spotify: The two volunteer trackers make the platform’s missing labels visible, but their own methods and the false-positive warning leave open questions about appeals and accountability. 404 Media’s report documents the gap.