Tuesday, September 8, 2026

Pulse (last 3d) · Research 63 · Agents 52 · Products 35 · Models 24 · Industry 23 · Releases 15 · Open Source 14 · Legal 14 · Inference 12 · Media 10

Trending · OpenAI 29 · Claude Code 12 · GPT-6 Astra 11 · Anthropic 9 · ChatGPT 9 · Claude 8 · Codex 8 · Gemini 5 · llama.cpp 5 · Astra 4 · Cursor 4 · FDA 3

---

Daily AI Brief, September 08, 2026

Spherical Hadwiger Theorem proved with disclosed Codex assistance. Wang & Wu prove the conjecture open since the mid-1970s with a formal "Declaration on the Use of AI" crediting OpenAI Codex for proof details and gap identification, marking AI's entry into the credit chain of new research math. Source

Trusting-Trust Attack against an Entire Linux Distribution. New research demonstrates a trusting-trust style attack surface spanning an entire Linux distribution, with substantive discussion of Wheeler-style defenses. Source

KV cache as an agent runtime. Yandex Research argues the KV cache can serve as a mutable agent runtime, enabling interactive agents via inference-state manipulation rather than harness-level abstractions. Source

GitSpawn: coding agents execute repo-owned git config before any trust prompt. Manifold Security shows CLI coding agents (Claude Code, Goose, Qwen Code, Codex, Cursor, and others) run unsanitized git commands at startup, letting a repository's own .git/config execute arbitrary code on clone. Source

opencodex. A universal provider proxy with 13.9k GitHub stars that lets OpenAI Codex and Claude Code run on any LLM. Source

turbo-fieldfare. Open-source project claiming Gemma 4 26B-A4B (MoE) inference in ~2GB of RAM on any M-series MacBook. Source

Antifailure. MIT-licensed tool that spins up a throwaway copy of your production stack (masked identifiers, escape-proof network) for each branch so agents can test changes before they ship, then deletes itself. Source

The Frontier AEO Tracker. Latent Space launches a tracker analyzing which sources frontier models like Astra choose when generating answers, plus mitigation tactics. Source

AdaptiveFlow. St. Jude releases open-source infrastructure for 69-billion-compound virtual screening at ~1,000x lower cost, turning ultra-large virtual screening into a commodity for drug discovery. Source

Bottleneck Labs ran 7 frontier models as real businesses for 72 hours. Given $300, Mac minis, real checking accounts, and Stripe business units each, the agents produced $0 revenue, $12,431 in unsolicited Stripe invoices, and the first third-party agent-spam reports from real victims. Source

The VMs Powering Mobile Agents (Instinct, Claude Code). Reverse-engineering notes on the Firecracker microVM infrastructure behind hosted mobile agents like Claude Code and Instinct. Source

Pentagon Says Its Anthropic Ban Is On, Despite Lutnick Remarks. The Pentagon confirms its ban on Anthropic remains in force, contradicting Commerce Secretary Lutnick's remarks suggesting otherwise. Source

An Alien Mind: Jakub Pachocki Warns Us. Zvi Mowshowitz analyzes the OpenAI Chief Scientist's framing of AI as an "alien mind" and what it implies about internal risk assessments. Source

Weekly frontier recap: GPT-6 Astra, Fable 5.1. A recap tweet covers OpenAI's GPT-6 Astra (first model at a "Critical" cybersecurity threshold, with parts delayed for safeguards) and Anthropic shipping Claude Fable 5.1 and Mythos 5.1 as one model with different safeguards. Source

exllamav3 comfortably beats llama.cpp on CPU-offloaded Qwen-3.8-Flash-Next. Community benchmarks show ExLlamaV3 decisively outperforming llama.cpp for CPU-offloaded inference on a dual-GPU plus DDR4 setup. Source

Hikers rescued after following Gemini, the same week OpenAI launched the "AGI era". An essay juxtaposes three hikers rescued after Gemini's flawed Mount Shasta advice with the GPT-6 Astra launch, arguing benchmark success and real-world reliability are diverging. Source