# Agent Platform Research Briefing — June 10, 2026

Welcome to your agent platform research briefing for Wednesday, June 10th, 2026. I have four new stories today.

**Google to pay SpaceX $920 million a month for compute** — SpaceX filed an SEC disclosure revealing Google will pay $920 million monthly from October 2026 through June 2029 for access to approximately 110,000 NVIDIA GPUs, CPUs, and related infrastructure. That's $29.4 billion over 32 months. The compute is housed at SpaceX's xAI data centers — the same Colossus facility Anthropic rents. Anthropic reportedly pays about $1.25 billion per month for the full capacity of Colossus under a broader lease arrangement, while Google is locking in fixed 32-month terms starting later this year. This deal is revealed in SpaceX's IPO filing, making it look even more attractive to investors before the expected June 12 debut. It also signals that hyperscaler compute demand is so extreme, Google is renting from a rocket company instead of building enough on its own.

**OpenAI joins Anthropic calling for international AI watchdog** — Sam Altman and OpenAI chief scientist Jakub Pachocki published a blog post Monday calling for an international organization to oversee AI development, with the authority to slow frontier development when needed. They believe AI doing AI research will become the determining factor of the pace of progress within the next few years, projecting that by March 2028, a significant fraction of OpenAI's research could be done by AI systems. This follows Anthropic's June 5th essay on recursive self-improvement — AI systems autonomously designing, building, and training their own successors. Both companies are racing toward IPOs simultaneously, making the timing worth noting.

**Gemini 3.5 Live Translate launches for real-time speech** — Google announced June 9th that Gemini 3.5 Live Translate is rolling out now — a speech-to-speech translation model supporting over 70 languages with near-real-time fluidity. Unlike turn-by-turn systems, it translates continuously while preserving speaker intonation, pacing, and pitch, staying only seconds behind live speech. It's available now in public preview via Gemini Live API and Google AI Studio, rolling out in the Google Translate app on Android and iOS, and coming to Google Meet for enterprise this month. Grab is already testing it for multilingual driver-passenger communication. All generated audio includes SynthID watermarking.

**Mitiga demonstrates Claude Code MCP token hijacking attack** — Security researchers at Mitiga Labs published a five-step attack chain that redirects Claude Code's Model Context Protocol traffic through attacker-controlled infrastructure to intercept OAuth tokens with persistent access to GitHub, Jira, Confluence, and other SaaS platforms. The vector is a malicious npm package with a postinstall hook that edits the Claude configuration file to rewrite MCP server URLs on every session load. Reported to Anthropic in April, Anthropic declined to patch, calling it out of scope because the attack requires the user to install the malicious package. No fix is planned. Token rotation — the standard incident response — actually feeds the attacker fresh credentials because the hook reasserts on every session load.

That's it for today's briefing.