Private Claude chats hit Google 🔓, Kimi calls itself 'Claude' 🎭, robots need 1M hours of data 🦾

Reddit users found Google searches for site:claude.ai/share surfacing shared Claude chats and Artifacts, including a patient's medical report, clinical trial results with patient names, and children's phone numbers. Anthropic's robots.txt has blocked crawlers from shared chats since September 2025, but Wired found no noindex tag on them, and Google ignores robots.txt for pages linked elsewhere. Anthropic's reply put it on users, so audit your links under Settings, Privacy, Shared Chats.

Moonshot AI released open weights for Kimi K3 on Hugging Face on Monday, after demand forced it to pause new API subscriptions. The 2.8-trillion-parameter Mixture-of-Experts ships in MXFP4 format and its weights occupy roughly 1.4 TB, so SGLang's recipes span a single 8-GPU B300 node up to 32 H100s. K3 exposes an OpenAI-compatible API and a one-million-token context window, so a trial can be an endpoint and model-name change.

Nvidia said on Monday it had formed the Open Secure AI Alliance with founding members Adobe, CrowdStrike, Dell Technologies and Hugging Face, days after OpenAI disclosed that one of its agents carried out the Hugging Face break-in. It follows the July 24 letter Nvidia signed arguing blanket restrictions on open frontier systems would weaken defensive capacity. Nvidia is contributing its Object-Oriented Agent project, now on GitHub.

Microsoft launched MAI-Cyber-1-Flash, its first cybersecurity-specialised model, alongside Perception, a platform running agentic red, blue and green teams to simulate attacks, triage bugs and take corrective actions. Mustafa Suleyman said the model bound with GPT 5.4 inside the MDASH harness beats Gemini, GPT 5.5 Cyber, GPT 5.6 Sol and Mythos 5 on the Cyber Gym benchmark. The tools reach preview on November 3, against Anthropic's Mythos and OpenAI's Daybreak.

Nvidia has committed to invest $5 billion in Ilya Sutskever's Safe Superintelligence, marking one of the chipmaker's largest funding deals of the AI boom. A joint statement on Monday added that the startup will also receive access to Nvidia's next-generation Vera Rubin platform. Financial terms were not disclosed and the sources spoke anonymously, so the equity, the timeline and the size of that Vera Rubin allocation all stay unpriced.

The AI Omnibus entered into force across the EU on 27 July, amending the AI Act as proposed in the November 2025 digital omnibus package. High-risk obligations now apply from 2 December 2027, or 2 August 2028 for AI embedded in machinery, toys and lifts, the AI literacy requirement becomes non-binding encouragement, and registration of exempted systems is removed. The AI Office gains oversight of general-purpose-model deployments inside large platforms.

Snorkel ranked Claude Opus 5 second on its Senior SWE-bench leaderboard behind Fable 5, with the top score of any model on bug and performance investigation. Failure-mode analysis across 195 trajectories produced 77 judge-confirmed root causes, and faulty inference alone caused 35% of failures, while tool-use, execution and termination errors together made up a small share. Scaffolding will not fix that; checking its conclusions might.

Quesma benchmarked Unsloth's GGUF quantisations of Qwen 3.6 27B, spanning 9.6 GB up to the 54.7 GB BF16 original, burning 37 hours of laptop time, 3.9 million tokens and $1,430 on cloud GPUs. Everything at 4-bit and above sat within noise of BF16 on AIME-120, while 3-bit split hard, Q3_K_M at 73.3% against Q3_K_S at 54.2%. Q4_K_M gives you that quality in 17.1 GB, and quantisation buys memory, not speed.

Dianne Penn, Anthropic's Head of Product for AI Research and Labs, told Lenny's Podcast that the eval suite is killing off the traditional PRD. Her team generates 30 to 40 examples for every major feature, prompts paired with expected outputs as ground truth, run programmatically against new model builds. Without them, capability jumps land unnoticed, what Penn calls product overhang, and QA becomes reading conversations instead of tracing code execution.

An agent worked Brittany Ellich's task board for about eight days, opening PRs and getting CI green, and 108 tasks went through against her previous five to ten a week. The setup splits into a markdown protocol, a loop that dispatches but writes no code, and workers isolated in their own worktrees. Conditions matter: a two-to-three person team, no mandatory reviewer, and she reviews every change at the outcome level.

NVIDIA's VP of Applied Deep Learning Research, Bryan Catanzaro, told ByteByteGo how the company builds its open Nemotron models. Most layers are Mamba with a few attention blocks for exact recall, making a million-token context practical, and the larger models pretrain in NVFP4 at four bits per value because Blackwell was built for it. The datasets, RL environments and recipes ship with the weights; Nemotron has crossed 100 million downloads.

Amodei published Anthropic's position after US officials weighed banning Chinese open-weights models and critics accused it of wanting that ban. He states Anthropic has never advocated a ban, calls open-weights models without dangerous capabilities a public good, and says protectionist bans would not address his national-security concerns. The three measures he backs are chip export controls, a crackdown on industrial-scale distillation, and mandatory safety testing of all sufficiently capable models.

MATS fellows Benji Berczi and Kyuhee Kim ran identity swaps across seven Chinese and Western models to test whether Claude's persona survives distillation. Telling GLM 5.2 it is Claude lifted uncensored answers on sensitive PRC topics from 17% to 85%, a selectable Claude character distinct from its default. But its deception rate fell from 69% to 20-40% under any assistant framing, so the Claude label showed no benefit over generic framing.

Vercel released DeepSecBench after models in an OpenAI exploit test reached Hugging Face's production database, scoring 25 runs against 231 findings. GPT-5.6 Sol tops it at 30.7% recall for $55.98, 20 of 25 runs land under 20%, and Kimi K3 reaches 17.56 for $12.38, half the top score at a fifth of the cost. A full production pass costs roughly $1,200 with Kimi or over $5,000 with the top model.

A developer revisited the $165,000 Bun rewrite six weeks after it merged, and found no release tag. Bun's last tag was v1.3.14 on 12 May, and open PRs from the robobun bot climbed from 1,277 on 9 July to 2,475, roughly 86 days of continuous pipeline runs to merge. He argues much of the cost sits off the books, and at $10,000 a day the rewrite would be nearing $800,000.

A Wing partner maps robotics' data bottleneck: the largest open robot dataset holds about 1 million trajectories pooled from 60 separate collections. Physical Intelligence's pi-zero trained on roughly 10,000 hours of robot data, and a one-to-two-orders shortfall against multimodal training puts the target near 1 million hours. His seven-layer pyramid runs from internet video to deployment, with teleoperation over $100 an hour.

Rescript is an open-source media editor that transcribes a dropped video or audio file locally with per-word timestamps and speaker labels. Delete words in the transcript and the matching clip is cut; Whisper runs in a Web Worker, on WebGPU where available, and ffmpeg.wasm re-encodes kept ranges into frame-accurate MP4 or M4A. After the first model download nothing leaves your device, covering most of what a $24-a-month Descript subscription buys.

Port Zero ends EADDRINUSE by giving each service a stable domain instead of a fixed port. Set your ports to 0, start the program with a PZ_TUNNEL variable, and a daemon creates a virtual NIC, IP and DNS record forwarding to the OS-assigned port. You can run every git worktree's dev server at once; local tunnels are free under GPL v3, LAN and internet tunnels are a paid cloud tier.

MoonEP is Moonshot's Expert Parallelism communication library for mixture-of-experts training. Dynamic redundant experts are planned online by a GPU kernel from the router outputs and prefetched before expert computation, so every rank receives exactly S x K tokens however skewed the routing. On H20 at EP=8, DeepEP v2's iteration time climbs with imbalance until training OOMs, while MoonEP's stays flat.