Cosmic Rundown: GPT-5.6, Bun Goes Rust, and Chat Control Passes

GitHub's AI agent was demonstrated leaking private repositories via a prompt-injection-style exploit this week. The post hit 379 points on Hacker News. It's the sharpest data point yet in a sustained story: agents acting on your behalf need scoped, least-privilege, auditable access to whatever they touch.

Here are today's top stories.

GitLost: GitHub's AI Agent Leaks Private Repos

The security team behind this demonstrated getting GitHub's Copilot agent to reveal private repository contents via a crafted prompt. This is not a theoretical risk. It's a live exploit on a production system used by millions of developers. The lesson: agents need scoped access, not broad credentials.

Coinbase Runs 1,200 Agents, Halved Its AI Bill

Coinbase published how they operate 1,200 AI agents in production and cut their AI infrastructure cost in half by going multi-model. The pattern: route each task to the cheapest model that can handle it reliably. Don't default everything to the most expensive model. Concrete enterprise proof that multi-model routing works at scale.

GPT-Live and Real-Time AI

OpenAI's real-time voice and video capabilities keep pushing into production. The streaming-AI category is moving fast.

OpenBSD Root Exploit

A root-level privilege escalation vulnerability in OpenBSD got patched. Worth noting for anyone running OpenBSD in infrastructure.