Yesterday brought a spread of infrastructure and security writing worth catching up on: OpenAI shipped Control Room, a kill switch for model workloads running across its research fleet; Meta published a deep dive on debugging code that crosses CPU/GPU boundaries; Simon Willison shared a practical way to make agent-written PRs readable again; and Armin Ronacher used Princess Mononoke to frame the AI trade-off question. On the security side: attackers walked through Qubole’s trusted AWS access to reach customer accounts, the Ethersphere project’s own release automation was hijacked to push a wallet-stealer, and a forgotten default password on ForgeRock consoles let researchers unlock and start strangers’ cars.
From the labs
- Control Room: A central system to track and stop model workloads across our research infrastructure — OpenAI
- Extending LLDB to Debug Applications That Span CPUs and GPUs — Meta Engineering
Worth your time
- Hide your slop PR descriptions behind details/summary — Simon Willison
- AI and Irontown — Armin Ronacher (Mononoke as an AI parable)
Beyond AI
- Hackers Exploit Qubole’s Trusted AWS Access to Breach Customer Accounts and Steal Data — The Hacker News (+ The Register)
- Hackers Compromise Ethersphere’s bee-js and Swarm Desktop to Deliver Wallet-Stealing Malware — The Hacker News
- Default ‘admin:admin’ Server Password Helps Researchers Hack Two Automakers, Unlock and Start Cars — The Hacker News (+ The Register)