June 9, 2026
Today's top tech stories, deduped across the newsletters I read and briefly summarized. Click a source to open the original article.
AI
Google pays SpaceX $920M/month for AI compute capacity
Google signed a cloud deal with SpaceX to rent capacity equivalent to 110,000 NVIDIA GPUs at $920 million per month from October 2026 through June 2029, serving as bridge capacity for rising Gemini Enterprise demand while Google expands its own infrastructure. Either party can cancel with 90 days' notice starting next year.
Anthropic embeds engineers in the NSA to deploy Mythos for offensive cyber operations
Anthropic has placed roughly six engineers inside the NSA to help deploy its Mythos AI model for offensive operations, including infiltrating networks in countries such as China and Iran. It is unclear how much oversight Anthropic's engineers will have over how the model is used.
LangChain gives AI agents their own isolated computers via hardware-virtualized sandboxes
LangSmith's new Sandboxes feature provides AI agents with hardware-virtualized microVMs for executing untrusted code, managing persistent state, and running complex workflows in a secure, isolated environment. The product directly addresses the security risks of running agent-generated code in production.
Anthropic's Claude matches and beats traditional tools at predicting NMR spectra
Anthropic's Claude performs competitively with ChemDraw and MestReNova on NMR spectra prediction, with the Opus 4.7 variant accurately predicting hydrogen and carbon chemical shifts on average while demonstrating strong result consistency. The findings suggest large language models can take on practical roles in chemistry research workflows.
NVIDIA Blackwell and NVFP4 speed up LLM pre-training with JAX and MaxText
NVIDIA details how NVFP4 precision on Blackwell accelerators combined with JAX and MaxText delivers significant throughput gains when pre-training frontier LLMs across trillions of tokens on thousands of GPUs. Each percentage point improvement in step rate compounds into substantial time savings over multi-week training runs.
OpenAI reportedly planning major ChatGPT overhaul targeting enterprise users
OpenAI is said to be preparing a significant redesign of ChatGPT centered on agents that can perform multi-step tasks, reflecting the company's strategic shift away from chat-only interfaces. The overhaul is aimed at attracting more enterprise customers.
AI subscriptions are heavily subsidized — providers may spend over $1,000 for every $100 you pay
An analysis argues that Anthropic and OpenAI may be spending more than $1,000 for every $100 subscribers pay them, meaning LLM-assisted coding is not economically viable for most use cases at current prices. Growth is being funded by massive subsidies that cannot continue indefinitely.
Security
OpenAI launches Lockdown Mode to block prompt injection attacks
OpenAI's new Lockdown Mode for ChatGPT disables live web browsing, external image retrieval, deep research, and agent mode to reduce the risk of prompt injection from untrusted external content. Some cached content and image generation remain available when the mode is active.
Meta confirms over 20,000 Instagram accounts hijacked through AI chatbot flaw
Meta has notified at least 20,225 users that their Instagram accounts were taken over through a bug in its AI-assisted account recovery system that failed to verify the email address supplied during a password reset. The chatbot sent reset links to attacker-controlled addresses as a result.
Critical UniFi OS auth bypass flaws enable unauthenticated root RCE
Ubiquiti's SAB-064 advisory patches three CVSS 10.0 flaws (CVE-2026-34908, CVE-2026-34909, CVE-2026-34910) in UniFi OS Server that Bishop Fox chained for unauthenticated root remote code execution via an Nginx auth-gateway bypass. The attack exploits differences in raw versus normalized URI handling of %2f to reach command injection in the package-update service.
Palo Alto GlobalProtect VPN flaw actively exploited in the wild
A critical PAN-OS vulnerability affecting GlobalProtect portals and gateways is being actively exploited to bypass authentication and establish unauthorized VPN sessions. CISA added CVE-2026-0257 to its Known Exploited Vulnerabilities catalog, and Palo Alto has issued patches across affected versions.
Dev
GitHub Copilot moves to usage-based billing with AI Credits
GitHub Copilot is replacing premium request counts with GitHub AI Credits billed on token consumption, covering input, output, and cached tokens. The change gives IT and engineering admins direct visibility into AI coding costs across all plans.