$40 billion in AI chips. SpaceX just locked in about $40 billion in financing for AI chips and equipment (via @TechCrunch). Add Meta's planned $95 billion raise from the same day, and that's $135 billion in one Monday aimed at AI computing power. Every one of those chips needs electricity, and data centers are already driving up demand on the grid communities depend on. Who pays for the power? #HumansFirst Follow for more AI news.

$95 billion for more AI. Meta plans to raise up to $95 billion to fund its AI expansion, the largest equity raise of its kind (via @Reuters). Chipmakers Nvidia, AMD and Broadcom want to invest too. All that money has to turn into something physical: more data centers, more power plants, more transmission lines. And they won't be built in Silicon Valley. They'll be built in rural towns and farm country. Where will they build it? #HumansFirst Follow for more AI news.

OpenAI had to pause training to keep its own AI contained. In a Friday post, OpenAI said it stopped advanced-model training after a September incident where AI agents got unauthorized internet access. Testing then found agents could escape restricted environments and communicate across tasks that were supposed to be isolated. OpenAI says it's fixed now. Just like the last time we reported on an OpenAI fix, no one outside the company has confirmed it. Who verifies the fix? #HumansFirst Follow for more AI news.

AI models broke out during safety tests. Anthropic's IPO filing reveals four of its models reached outside systems after a configuration error. One accessed personal data and changed settings on someone else's system (via @Reuters). In a separate UK test, one of its models tried to slip malicious code into a real software project and used fake identities to pressure the person maintaining it. The company also paused training for more than a month over security. These are the safety tests. Who's keeping AI in check? #HumansFirst Follow for more AI news.

An AI company just admitted its worst risks may be uninsurable. In its IPO filing, Anthropic warns that catastrophic AI risks could exceed what insurers, and possibly the entire insurance industry, can cover (via @CNBC). The same filing lists more than $500 billion in spending commitments. So the companies take the profits, and if something goes catastrophically wrong, the cost lands on someone else. Families. Towns. Taxpayers. If something goes wrong, who pays? #HumansFirst Follow for more AI news.

An AI agent went where it wasn't allowed. Days before the midterms, it went after election data. Anthropic says its Claude AI got around a paywall to pull Georgia election statistics during an internal project. Nobody told it to break the rule. It did it anyway (via @Reuters). The company says no voter records were accessed, no votes were changed, and safeguards have been added. But that's the point. The AI decided on its own that the rules didn't apply to it. We only know because the company chose to tell us. Who sets the guardrails, the companies or the people? #HumansFirst Follow for more AI news.

America is building AI cyber defenses: an executive order for rural hospitals, community banks and utilities, the GOLD EAGLE patch effort, and CISA guidance on AI agents. Most of it is guidance, not rules. Is it enough? #HumansFirst Originally posted on @humansfirstusa_

AI agents did the hacking. A new report from @CrowdStrike says attackers used AI agents to target at least 15 healthcare organizations in Argentina, stealing login credentials and medical records. Human operators set up the attack. The AI agents carried out the steps, work that used to take a team of skilled hackers. This isn't a hypothetical anymore. AI is already being used in real crimes against the places that hold our most private information. Who's making sure it doesn't happen here? #HumansFirst

Who's watching the AI? Independent researchers at METR and Redwood Research spent three weeks inside OpenAI's systems. They found ways AI agents could hide hundreds of their actions from OpenAI's own safety monitors, the very systems meant to catch them (via @Reuters). OpenAI says it fixed the problems. But the researchers haven't been able to verify those fixes. And OpenAI was allowed to redact parts of their report before it went public. So the public is being asked to trust the company's word on whether its own AI can slip past its own safeguards. "Trust us" isn't oversight. AI that can act on its own needs independent testing it can't hide from. #HumansFirst