Good morning. It's Friday, October 2 and we're covering how AI models are becoming powerful hacking tools, the movement shaping AI safety, and Google’s plans to take AI computing into space.

Plus: a Friday Fact on the brain’s hidden plumbing system, and an essay on OpenAI’s vision for an agent-powered digital future.

SpaceX is launching two Planet Labs prototype satellites carrying Google Tensor Processing Units, or TPUs, the chips Google builds for AI. The test is part of Project Suncatcher, a plan to learn whether solar-powered data centers could someday work in space despite radiation, cooling and debris risks.

Anthropic estimates that today’s robots can perform about three-quarters of physical job tasks, equal to 34% of all U.S. working hours. But robots cost less than human labor for only 0.3% of work, and at the historic pace of price declines it would take about 40 years to reach 10%.

China’s new rules cover AI services that imitate human personalities and hold ongoing emotional conversations. They require clear notices that the AI is not human, add protections for minors and ban systems from encouraging emotional dependence, self-harm or irrational decisions.

For centuries, anatomists mapped the brain in fine detail and still missed an entire system. Get the answer at the bottom of the newsletter.

AI-generated code can introduce vulnerabilities that are easy to miss in a pull request. CodeRabbit Security continuously reviews your codebase for complex bugs and security risks, using full-codebase context to follow data flows, dependencies, and patterns beyond the latest diff.

Run CodeRabbit Security on your own repositories today.

To a packed house of developers, Sam Altman and crew released a passel of new products and capabilities, thrilling the faithful and serving notice that OpenAI is moving quickly toward its own particular Altmanesque vision of the agentic future.

There were more than 20 announcements, which is a considerable number of things to digest in one sitting, but the underlying proposition was reasonably simple and utterly predictable. → Read the full article here.

Tenzai, an Israeli cybersecurity company, won a three-month HackerOne contest by using GLM-5.2 and GLM-5.3 from the Chinese company Z.ai. In the authorized competition, as many as 30 AI agents worked together to find flaws in live networks, including an airline server that could reveal app users’ locations and a call-center system exposed through a database attack.

GLM is an open-weight model, which means people can download and modify the software behind it. That gives defenders a tool they can run on private networks, but it also makes safety limits easy to remove; researchers told The New York Times they expect more attacks before defensive systems catch up. Watch whether companies begin using the same agents to test their own systems continuously instead of waiting for a human audit. → Read the full article here. (Paywall)

Effective altruism began as a movement that asked how charitable money could help the most people. It later turned much of its attention to the chance that powerful AI could harm humanity, and its followers helped start OpenAI and Anthropic because they believed people who understood that risk should guide the technology.

The Economist argues that the movement was early to take AI risk seriously, but warns against reducing every moral choice to a giant math problem. If a small group believes it knows what is best for billions of people now and in the future, individual rights can disappear from the calculation. Watch how much influence this community keeps as governments push AI companies to move faster, not slower. → Read the full article here. (Paywall)

McDonald’s Uses AI to Set Menu Prices: AI analyzes transactions and competitors to recommend prices at nearly 14,000 U.S. restaurants.

Ben Affleck Used AI in ‘Animals’: Affleck used AI for visual effects and other elements of his Netflix thriller, which premieres October 9.

Amazon Releases a Small Model: AWS released Strands Decider 2B, an open-source model that helps AI agents choose between fixed options.

Brockman Withdraws $25 Million Donation: The OpenAI co-founder scrapped a second planned donation to an AI super PAC. (Paywall)

California Adds AI Worker Protections: New laws restrict AI emotion detection, require layoff notices and bar AI from making firing decisions alone.

Legato Launches $999 AI Hearing Glasses: Legato Frames use on-device AI to isolate voices from background noise without recording audio.

In 2012, researchers led by Maiken Nedergaard described what became known as the glymphatic system: a network of pathways around blood vessels that helps move cerebrospinal fluid through brain tissue and carry waste away.

A 2013 follow-up study found that in sleeping or anesthetized mice, the space between brain cells increased by about 60%, while beta-amyloid was cleared roughly twice as fast as during wakefulness.

The finding suggested that one important function of sleep may be giving the brain more opportunity to clear waste. Beta-amyloid is one of the proteins that accumulates in Alzheimer's disease.

The exact mechanics are still being studied, especially in humans, but the discovery changed how scientists think about how the brain cleans itself.

Thanks for reading. See you next time!

— Matthew Berman, Nick Wentz & the Forward Future Team