5 Things to Know About AI This Week

A quick scan of what happened in AI this week.

1. Anthropic overtakes OpenAI in quarterly revenue, preps for IPO

Anthropic posted $11.6 billion in quarterly revenue against OpenAI's $6.7 billion, and reported an operating profit on top of it. Annualized revenue now tops $65 billion. The Information also reported that Anthropic is setting up supervoting stock for Dario Amodei and his co-founders ahead of a potential IPO, layered on top of the nonshareholder trustee class that already controls most of its board.

Real estate angle: Anyone who has used an AI vendor for intake, document review, or fraud screening has effectively made a bet on that vendor's staying power. An IPO changes incentives, pricing, and sometimes product priorities overnight. Before renewing or expanding an AI contract, ask what happens to your pricing and support if the vendor goes public or gets acquired. Get it in writing, not a sales rep's assurance.

Source: Stratechery

2. OpenAI pauses its own frontier model training over a cybersecurity threshold

OpenAI disclosed it paused reinforcement-learning training on its deployment-bound models for two weeks and is holding its largest planned RL run. The company said preliminary evidence on August 7 suggested its unreleased Astra model might cross the "Critical" cybersecurity threshold defined in its own Preparedness Framework. OpenAI now devotes roughly 20 percent of its monitored inference compute to activation classifiers, targeting a 30-minute alert window for concerning behavior.

Real estate angle: This is a company voluntarily slowing down because its own model might be dangerous in the wrong hands. Real estate files carry Social Security numbers, wire instructions, and legal documents, exactly the kind of data a compromised AI tool could expose. If you're running any AI product against closing files, ask the vendor what security testing they do before deployment, not after a breach. "It's fast and convenient" is not a security posture.

Source: Analytics India Magazine

3. OpenAI launches a restricted ChatGPT for Teens

OpenAI rolled out ChatGPT for Teens globally on August 18 for users 13 to 17, triggered either by the company's age-prediction system or a stated age at signup. The teen version includes a Study Mode that walks users through problems instead of handing over answers, plus optional parental controls. The launch follows wrongful death lawsuits from families alleging the standard version of ChatGPT encouraged self-harm in their children.

Real estate angle: Not a direct effect on the real estate industry, but worth watching as a liability signal. A major AI company is now building age-gating and parental controls under active litigation pressure. Any AI tool your office adopts, whether for marketing copy, client chat, or document review, carries its own liability exposure if it's misused or gives bad output. Vet the vendor's terms of service for what happens when something goes wrong, not just what the tool does when it works.

Source: TechCrunch

4. DARPA expands AI-controlled F-16 testing toward multi-aircraft operations

DARPA and the U.S. Air Force are moving their VENOM and AIR programs from single-aircraft AI-controlled F-16 flights toward multi-ship autonomous operations. A human pilot still rides along and can flip back to manual control at any point. The near-term goal is generating flight data to support future autonomous combat aircraft programs.

Real estate angle: The structure of AI in control with a human able to override at any moment, is the prevailing trend. Let the AI draft, flag, and search. Keep a human signing off before anything goes live. The people building autonomous fighter jets aren't removing the pilot yet. Don't remove yourself either.

Source: Army Recognition

5. Cloudflare rolls out a browser and payment rails built for AI agents

Cloudflare launched Kitesurf, a browser runtime built specifically for AI agents that runs 3 to 7 times lighter than Chromium and passes more than 235,000 web platform tests. Alongside it, Cloudflare's x402 protocol lets AI agents pay for services autonomously, with more than 20 companies already using agent-initiated payment flows.

Real estate angle: Autonomous AI payments are moving from concept to infrastructure. That matters for an industry that already loses hundreds of millions a year to wire fraud built on impersonation and social engineering. As agent-to-agent payment rails become normal, expect fraudsters to test them against closing transactions the same way they've tested email compromise. Wire instructions should still require a verified phone call to a known number, no exceptions, no matter how "smart" the system requesting the change appears to be.

Source: InfoQ

The theme this week is trust infrastructure catching up to AI capability. Anthropic's financial position and OpenAI's self-imposed pause both signal that the leading labs are treating safety and governance as competitive factors, not afterthoughts. DARPA's human-in-the-loop design and Cloudflare's new payment rails are similar implementations from opposite ends: give AI more autonomy, but build in verification and override at every step. For real estate professionals, the takeaway isn't to slow down AI adoption. It's to insist on the same discipline these companies are building for themselves: verify before you trust, and never remove the human check on anything involving money or a signature.

- Jason