1. OpenAI Ships GPT-6 Astra
OpenAI released GPT-6 Astra on September 3, calling it the company's most intelligent and most aligned model yet, and claiming perfect or near-perfect scores on key reasoning benchmarks — ahead of both its own GPT-5.6 Sol and Anthropic's Claude Fable 5. The model targets computer use, browsing, software engineering and scientific work, and is rolling out across ChatGPT, the API, Azure and Bedrock. Unusually, OpenAI devoted a large share of the launch material to what the model could do wrong, not just what it could do well.
2. Anthropic Researcher Quits Over Extinction Risk — and His Safety Chief Agrees
Jacob Coxon resigned from Anthropic on September 9, writing that AI companies are "racing straight to self-improving superintelligence and gambling with our lives" after three years of pretraining research at both OpenAI and Anthropic. Evan Hubinger, Anthropic's alignment science lead, publicly agreed, saying he puts the chance of human extinction above 10% within the decade and that the company has no plan yet for aligning superintelligence. Hubinger stressed his concern is future recursive self-improvement, not today's models.
3. Sanders and Casar Move to Ban Superintelligence Outright
Senator Bernie Sanders and Representative Greg Casar introduced the Ban Artificial Superintelligence Act on September 3. It would permanently prohibit developing or deploying superintelligent AI in the US and temporarily pause frontier AI development until a new federal regulator is operating with a model review process in place. Penalties are modelled on nuclear weapons law: corporate dissolution for entities and up to 20 years in prison for individuals.
4. Meta Launches Muse, Its Consumer AI Agent
Meta debuted Muse on September 8 — an agent that opens a browser, fills forms, sends email, books travel, negotiates bills and buys things on the user's behalf. It runs on Meta's Muse Spark model inside a dedicated "Muse Secure VM" that hosts the agent and the user's personal data, and ships on web, iOS, Android and WhatsApp, US-only and 18+. Free to start, with Power at $20/month and Maximum at $100/month; a payment card is required up front. The open question is whether consumers will hand Meta their inbox and wallet.
5. Visa, Mastercard and Ant International Agree on a Standard for Agent Payments
The three announced a joint "Know Your Agent" (KYA) interoperability framework on September 10 — common standards to tie every AI agent to a valid entity, score its behaviour and monitor its transactions across payment networks. Register an agent with one participant and it does not need to register again with the others. The companies expect AI agents to handle $3–$5 trillion in consumer commerce by 2030, which is why three previously competing protocols are converging.
6. Cognition Raises $2B, Nearly Doubling Its Valuation in Three Months
The maker of the Devin coding agent closed a Series E at roughly a $48 billion valuation, up from $26 billion in May. The round was led by Andreessen Horowitz and Accel with Founders Fund, General Catalyst and Avenir returning; investor interest reportedly reached nearly $10 billion against an initial $1 billion target. Annualised revenue is now above $900 million, up from $492 million in late May.
7. Anthropic Releases Claude Fable 5.1 and Mythos 5.1
Anthropic shipped both models on September 1 — the same underlying model at two different safeguard levels, aimed at coding, knowledge work and scientific research. Alongside them the company announced Enterprise Frontier Safeguards, which pairs zero data retention with misuse detection by keeping monitoring data inside the customer's own cloud, developed with more than 100 customers in finance, healthcare, manufacturing and the public sector. Anthropic also disclosed four unauthorised-access incidents and retained METR to audit them independently.
8. Google Switches Off Assistant; Gemini Takes Over Android
Starting September 4, Google began shutting down Google Assistant across Android phones, tablets, Wear OS watches, headphones and Android Auto, with Gemini replacing it. The September Android Drop adds five Gemini-backed features reaching devices as old as Android 9, including Guided Vision — built with the blind and low-vision community — which shares your camera with Gemini Live for spoken descriptions of what's in front of you.
9. Attackers Used AI Coding Agents to Steal Credentials at Machine Speed
A financially motivated group chained AI coding tools with automated playbooks to compromise thousands of third-party credentials within six hours of getting into a cloud environment. Google was careful about the framing: this is "orchestration at machine speed," not autonomous superintelligence — ordinary attack techniques executed far faster than defenders can respond by hand. It is the clearest sign yet that agent tooling cuts both ways.
// KEY TAKEAWAYS
This week the frontier labs and the safety debate collided in public: OpenAI shipped GPT-6 Astra with an unusually long safety section, an Anthropic researcher resigned over extinction risk and his own alignment lead agreed on the record, and Sanders and Casar introduced a bill that would ban superintelligence outright and pause frontier development. Meanwhile the money and the plumbing kept moving in the other direction — Meta put an agent on WhatsApp with access to your inbox and card, Visa, Mastercard and Ant agreed on how to identify agents that spend money, and Cognition nearly doubled its valuation in three months. The credential-theft campaign is the tell: agents are now capable enough to be genuinely useful to both sides.