🧠 AGI is finally here

GPT-6 has landed

Happy Tuesday, legends. Welcome back to another edition of The Frontier, our weekly newsletter covering the best new AI launches on Product Hunt.

---

WHAT'S HOT

🧮 Fermat’s Last Theorem now has a computer-checked proof, and Claude wrote it. Eleven days, 13 million lines of Lean and 30,300 theorems, published by Anthropic on 4 September. (Anthropic)

💸 Mistral raised €3 billion, the largest equity round a European technology company has ever closed. Samsung Electronics led it at a post-money valuation above €21 billion, three years after the company started. (Mistral)

🔓 CISA says attackers are already inside LiteLLM gateways. CVE-2026-59822 opens an authenticated MCP session to anyone sending any bearer token, and it was being chained to drop crypto miners before CISA listed it on 3 September. (The Hacker News)

📱 A 2.5-billion-parameter model is outscoring 4-billion-parameter ones. OpenBMB’s MiniCPM5-2B averages 53.9 across 34 benchmarks against Qwen3.5-4B’s 51.1, and scores 97.1 on tool use where the models it is measured against manage 6.8 to 20.8. (MarkTechPost)

🧠 OpenAI shipped GPT-6 Astra and admitted it can no longer reliably watch the model think. The system card records a “substantial decrease” in chain-of-thought monitorability, and the line “If the model were to try to sandbag covertly, we would likely be unable to catch it.” (Transformer)

---

The week in products

 

 

---

FROM THE FORUMS

Scope the verb, not the agent

An agent with your permissions finishes the job. An agent with fewer stops halfway through something you told it to do. Shivangi Tripathi (@roguetink) asked which way people go, and nobody picked narrow scopes.

Siarhei Hamanovich (@siarheihamanovich) puts the guardrail in the verb instead, so his text-to-speech action won’t take a string, it takes a row id and reads the text server-side. James Cummings (@jamescummings) gates on reversibility rather than resources, a short list of money, deletion and anything that leaves the building. Aleksandra Zhdanova (@aleksandrazhd) makes every write to her MCP server two calls, a dry run returning a token bound to those exact parameters and a second call that fails unless they match: “Narrow scopes didn’t get us there; the two-step did.”

Good thread if your agent has ever stopped dead halfway through a task because you scoped it properly.