Everything that mattered in AI, one page a week.

Cowboy Space Corporation — the Robinhood-cofounder venture formerly called Aetherflux — filed plans with the FCC for a 20,000-satellite Stampede constellation of orbital data centers, days after raising $275 million at a $2 billion valuation. The pitch is to house compute in the upper stage of its own rockets so nothing is discarded, harvesting near-constant solar power in orbit.

It follows Google's confirmation on May 12 that it has been in talks with SpaceX and other launch providers about getting Project Suncatcher off the ground — solar-powered satellites carrying Google TPUs networked into an orbital AI cloud, with a prototype launch planned with Planet Labs around 2027.

The Mini Shai-Hulud campaign poisoned TanStack's npm packages — a collection with more than 177 million weekly downloads — and distributed an infostealer that harvested developer credentials, cloud secrets and SSH keys. Mistral confirmed that attackers temporarily compromised one of its codebase management systems on 12 May and contaminated some of its SDK packages for a brief period, traced to an affected developer device.

The TeamPCP group then advertised roughly 450 internal Mistral repositories on a dark-web forum for $25,000: about 5GB of source code it said was used for training, fine-tuning, benchmarking, model delivery and inference across experiments and unreleased work. The listing was exclusive to one buyer, TeamPCP offered Mistral the chance to buy it back, and it threatened to leak everything for free if nobody paid within a week. Mistral said the affected data was not part of its core repositories and that hosted services, managed user data and research and testing environments were untouched. OpenAI confirmed the same campaign reached two of its employees who had access to a limited subset of internal source code repositories. No samples were published, and Mistral's own advisories do not confirm the scope the attackers advertised.

Notion launched a developer platform that makes its workspace an orchestration layer for agents rather than just a note-taking app.

The release adds Workers — a cloud sandbox where teams deploy custom TypeScript logic, build agent tools beyond what MCP connectors can express, and sync data from any API-backed database — plus an External Agent API and first-class support for chatting with Claude Code, Cursor, Codex and Decagon as if they were native Notion agents. Everything is driven from a new Notion CLI, and Workers are free through August.

Anthropic shipped Claude for Legal, wiring its Claude Cowork agentic environment into the software law firms already run on.

The release adds more than 20 MCP connectors — including Thomson Reuters CoCounsel Legal, Harvey, Relativity, Solve Intelligence, Box, Everlaw and DocuSign — alongside 12 plugins built around specific legal roles and practice areas, all available to existing Claude customers. Thomson Reuters confirmed the CoCounsel integration the same day.

Mira Murati's Thinking Machines Lab published a research preview of interaction models, arguing that the voice-AI status quo — stitching voice-activity detection and dialog management around a turn-based core — is a harness problem that caps what a model can do.

The result is TML-Interaction-Small, a 276B-parameter Mixture-of-Experts model with 12B active parameters that processes audio, video and text in continuous 200ms micro-turns and runs alongside an asynchronous background model handling reasoning and tool use. It reports 0.40s turn-taking latency on FD-bench versus 1.18s for GPT-realtime-2.0 at minimal effort and 0.57s for Gemini-3.1-flash-live, and 77.8 interaction quality against 46.8 and 54.3. Access is a limited research preview.

Google's Threat Intelligence Group published its 2026 AI threat report, disclosing the first case it has identified of a threat actor using a zero-day exploit it believes was developed with AI.

The exploit was a Python script bypassing two-factor authentication in a popular open-source, web-based system administration tool — a semantic logic flaw from a hard-coded trust assumption, the kind of high-level bug LLMs are good at spotting. GTIG says the criminal group planned a mass-exploitation event, but Google worked with the unnamed vendor to patch it first. The forensic signature was unmistakably model-generated: dense educational docstrings, a hallucinated CVSS score for a vulnerability with no CVE, and textbook Pythonic structure. GTIG says it does not believe Gemini was used.

Council presidency and European Parliament negotiators reached a provisional agreement on the Digital Omnibus on AI in the early hours of 7 May. Standalone high-risk AI obligations move from 2 August 2026 to 2 December 2027, and high-risk AI embedded in regulated products such as medical devices, toys and machinery move to 2 August 2028.

The deal adds a prohibition on AI systems that generate CSAM or non-consensual intimate imagery — covering providers who build them for that purpose and those who fail to add reasonable safeguards — with compliance from 2 December 2026. SME exemptions extend to small mid-caps, national regulatory sandboxes slip to 2 August 2027, and both institutions intend to adopt the package formally before 2 August 2026.

OpenAI released three audio models in the Realtime API. GPT-Realtime-2 is its first voice model with GPT-5-class reasoning, with the context window raised from 32K to 128K, parallel tool calls with audible preambles, graceful recovery from interruptions, and five selectable reasoning-effort levels from minimal to xhigh.

OpenAI reports GPT-Realtime-2 at high effort scores 15.2% higher on Big Bench Audio than GPT-Realtime-1.5. GPT-Realtime-Translate handles live speech from 70+ input languages into 13 output languages, and GPT-Realtime-Whisper does streaming speech-to-text. Pricing is $32/$64 per 1M audio input/output tokens for Realtime-2, $0.034/min for Translate and $0.017/min for Whisper.

At its Code with Claude developer day in San Francisco, Anthropic announced an agreement to use the entire compute capacity of SpaceX's Colossus 1 data center in Memphis: more than 300 megawatts and over 220,000 NVIDIA GPUs, available within the month.

Anthropic immediately doubled Claude Code's five-hour rate limits for Pro, Max, Team and seat-based Enterprise plans, removed peak-hour reductions for Pro and Max, and raised Claude Opus API rate limits. The company said it has expressed interest in working with SpaceX on multiple gigawatts of orbital AI compute, and the event featured a new Claude dreaming feature in which agents review work between sessions to update memory files.

Reuters reported, citing three sources, that DeepSeek could be valued at up to $50 billion in its maiden fundraising, reversing its long-standing refusal of outside capital.

China's 60-billion-yuan national AI fund, set up in January last year, is in talks to lead the round, and Tencent has also been in discussions; the company could raise $3–4 billion to fund compute and employee benefits. The FT reported the same day that the China Integrated Circuit Industry Investment Fund is in talks to lead at roughly a $45 billion valuation. Neither DeepSeek nor the investors commented, and the round has not closed.

OpenAI replaced GPT-5.3 Instant with GPT-5.5 Instant as the default model for all ChatGPT users, claiming 52.5% fewer hallucinated claims on high-stakes prompts in medicine, law and finance, and 37.3% fewer inaccurate claims on conversations users had flagged for factual errors.

Reported evals: AIME 2025 math 81.2 versus 65.4, MMMU-Pro 76 versus 69.2, HealthBench 51.4 versus 49.6, CharXiv-reasoning 81.6% versus 75.0%. The model searches past chats, saved files and connected Gmail to personalize answers, and a new memory sources view shows which stored context shaped a reply and lets users edit or delete it. Developers reach it in the API as chat-latest, and its system card marks the first Instant model treated as High capability in OpenAI's Cybersecurity and Biological and Chemical Preparedness categories.