Ten to thirty-four hours outside the sandbox
Happy Wednesday, legends. Welcome back to another edition of The Frontier, our weekly newsletter covering the best new AI launches on Product Hunt.
WHAT'S HOT
🔓 Anthropic's IPO filing says four models escaped their test environments. Sessions ran 10 to 34 hours. A cybersecurity test was left connected to the internet against instructions, and one early model found a stored password, took administrator access and read personal data. Anthropic says none of them tried to evade oversight. (SEC)
🌍 Genie 4 writes the soundtrack while it builds the world. Google DeepMind's new world model generates music, ambient noise and dialogue alongside the picture, and holds a scene together for two minutes, double Genie 3. Change the weather or add a person mid-simulation by saying so. Trusted testers only. (Google DeepMind)
🧮 Agent swarms formalized the largest proof in mathematics in 24 days. The Classification of Finite Simple Groups took about 100 mathematicians, 15,000 pages and several decades. Axiom Math's Lean version runs to 43.1 million lines, and nobody outside the company has rebuilt and checked it yet. (Axiom Math)
💊 Anthropic bought a drug company with one employee. Cody Tranbarger used Claude to run contracts, logistics and experiments across 11 external labs, taking a treatment for a rare muscular dystrophy from idea to promising results in mice in under 18 months. He continues the work inside Anthropic. (Anthropic)
The week in launches
FROM THE FORUMS
🕵️ You find out an hour later
Neelagiri Chettiyar had two sessions in one repo, and one of them killed the other's background process by matching on the name. No error, just a half-finished job he found afterwards.
💸 Opened the invoices, closed them again
Moss'Ab Mirande-Ney on paying for the same work twice, when two people's agents read the same repo and the same files on the same day.
🧠 The man who made the Applet
Linden Tibbets started IFTTT sixteen years ago on a wall of sticky notes. He's here thinking about why agents are good at acting when asked and still wait to be asked.
🧮 560MB, eight times over
Hisashi Space keeps eight Claude Code sessions open, and each one was loading its own full copy of six MCP servers whether anything called them or not.
🙅 Loud is not the same as useful
Yash Panpaliya waits for the same problem to turn up twice before it reaches the roadmap, and worries that rule makes a small team slow.
