AI in 15 — April 15, 2026

OpenAI just raised a hundred and twenty-two billion dollars, and its own investors are already asking if they overpaid. When the people writing the checks start questioning the valuation, you know something interesting is happening.

Welcome to AI in 15 for Wednesday, April 15, 2026. I'm Kate, your host.

And I'm Marcus, your co-host.

Happy Wednesday, Marcus. Packed show today. OpenAI's eight hundred and fifty-two billion dollar valuation is under fire from its own backers. OpenAI and Anthropic are in a full-blown cybersecurity arms race. Claude Code gets background agents that work while you sleep. A Swiss doctor vibe-coded a patient management system and exposed everything. Google turns Chrome into an AI app platform. And a new research paper could change how every AI model generates text. Let's get into it.

OpenAI investors question whether the biggest funding round in history was worth it.

The AI cybersecurity arms race heats up with dueling strategies.

And vibe coding meets healthcare with terrifying results.

Marcus, the Financial Times is reporting that investors in OpenAI's record funding round are publicly questioning the company's direction. What's going on?

The core tension is strategic focus. OpenAI has ChatGPT with a billion users, the consumer AI category it essentially created. But Sam Altman is pivoting hard toward enterprise, chasing Anthropic's territory. One early backer told the FT, and I'm quoting here, "You have ChatGPT, a billion-user business growing fifty to a hundred percent a year. What are you doing talking about enterprise and code? It's a deeply unfocused company."

And the numbers behind Anthropic's growth are eye-popping.

Anthropic's annualized run rate surged from roughly nine billion at the end of 2025 to thirty billion by end of March 2026. That's largely driven by developer tools like Claude Code. OpenAI sits at about twenty-five billion annualized, with enterprise now forty percent of revenue. So Anthropic has actually overtaken OpenAI on raw revenue run rate while being valued at less than half. Three hundred and eighty billion versus eight hundred and fifty-two billion.

So investors are looking at that math and getting nervous.

The math is brutal, Kate. For investors in the recent round to break even, OpenAI needs to reach a one-point-two trillion dollar IPO valuation, potentially as early as late 2026. Meanwhile, OpenAI's chief revenue officer accused Anthropic of overstating revenue by roughly eight billion dollars, which Anthropic disputes. And secondary market data shows some buyers now preferring Anthropic shares over OpenAI.

OpenAI also shut down Sora after burning an estimated fifteen million dollars a day in compute. That can't help investor confidence.

Active users collapsed to under five hundred thousand. Six months after launch. And then there's the TBPN acquisition that one investor called a distraction. Look, OpenAI still has enormous assets. A billion users, massive compute capacity, GPT-5.4 with a million-token context window. But when you're valued at eight hundred and fifty-two billion, everything has to work. There's no room for expensive experiments that don't pan out.

As we discussed Monday, the market is demanding proof now. OpenAI is exhibit A.

From valuations to vulnerabilities. We've been covering Anthropic's Mythos and Project Glasswing all week. Now OpenAI has fired back with GPT-5.4-Cyber. Marcus, what's the play here?

OpenAI launched a fine-tuned version of GPT-5.4 specifically for defensive cybersecurity. It's what they call cyber-permissive, meaning lowered refusal boundaries for legitimate security work. The standout capability is binary reverse engineering, analyzing compiled software without source code. Access goes through a tiered verification system for security professionals.

And the strategic difference between the two companies is fascinating.

It really is. Anthropic restricted Mythos access to about forty handpicked organizations through Project Glasswing. Apple, Amazon, Microsoft, CrowdStrike, the Linux Foundation. OpenAI is betting on broader distribution to thousands of individual defenders and hundreds of security teams. Two completely different philosophies for handling the same dual-use problem.

Anthropic's approach is elite and controlled. OpenAI's is democratized.

And both have risks. Anthropic's model means most organizations don't get access to the best defensive tools. OpenAI's model means more potential for misuse despite the verification system. OpenAI's Codex Security has already contributed to fixes for over three thousand critical vulnerabilities, and their capture-the-flag scores jumped from twenty-seven percent to seventy-six percent in about three months. So the capability is real.

As we covered Sunday with the AISLE counter-study, the moat isn't the model, it's the system. Both companies seem to agree on that now, just building different systems.

Exactly right. And the urgency is real. Mythos found thousands of high-severity zero-days across every major operating system. On Firefox alone, a hundred and eighty-one successful exploits. Vulnerability discovery is now outpacing patching. That's not a future problem, Kate. That's today.

Sticking with Anthropic. Claude Code just got a major new feature called Routines. This feels like a significant shift in what a coding tool can be.

Routines are automations that run on Anthropic's cloud infrastructure, not your local machine. You package a prompt, repositories, and connectors into a saved configuration. Then you can trigger it three ways: on a schedule, via API, or through GitHub events like pull requests and pushes. The key thing is these keep running when you close your laptop.

So practical examples?

Nightly issue triage. The routine reads your GitHub issues, applies labels, assigns owners, posts a Slack summary. Or alert-based triage where monitoring tools trigger a routine when error thresholds spike, and it correlates stack traces with recent commits. Automated PR reviews. The use cases are broad.

This is moving from tool to teammate.

That's exactly the framing. And combined with Anthropic's thirty billion dollar run rate, it explains the strategy. Make Claude Code indispensable, not just when you're at your desk but around the clock. Pro users get five runs daily, Max gets fifteen, Team and Enterprise get twenty-five. The Hacker News discussion was mixed, some loved it, others worried about usage limits and feature sprawl. They also shipped a redesigned desktop app with parallel sessions and an integrated terminal.

Now for a story that should make everyone uncomfortable. A blog post went viral about a Swiss medical professional who vibe-coded an entire patient management system. Marcus, how bad is it?

About as bad as it gets. This person watched a video about building software with AI, then proceeded to create a patient management system. They imported all existing patient data, published it to a US server without a data processing agreement, and added a feature to record patient conversations and send audio to two AI services for transcription. All without patient consent.

And the security?

The entire application was a single HTML file. The backend had zero access control. Data was literally one curl command away from anyone who looked. All access control logic lived in client-side JavaScript. Everything unencrypted. Completely exposed for several days before discovery. Potential violations of Switzerland's data protection law and professional secrecy laws.

This is the nightmare scenario we've been warning about.

And it's not isolated. Georgia Tech's Vibe Security Radar tracked thirty-five CVEs attributed to AI-generated code in March alone, up from six in January. Veracode found forty-five percent of AI-generated code contains security flaws. Apple is now booting vibe-coding apps like Replit and Vibecode from the App Store over concerns about malicious code. The gap between "I can build this" and "I should deploy this" is becoming a public safety issue.

When the tool is powerful enough for anyone to build an app in hours but the builder doesn't understand authentication, encryption, or data protection laws, this is what happens.

And healthcare is the worst possible domain for this to happen in. Patient data is among the most sensitive information that exists. The vibe coding movement needs guardrails, and frankly, the AI tools themselves should be flagging when generated code has no access controls or is storing sensitive data unencrypted. That's a solvable problem.

Google announced Skills in Chrome. Basically turning saved Gemini prompts into reusable one-click workflows in the browser. Marcus, is this meaningful or just a feature?

It's strategic. You save a prompt as a Skill, then invoke it later with a slash command or click. Skills work across different web pages and tabs. Calculate protein macros on any recipe page. Generate spec comparisons across shopping tabs. Scan documents for key points. Rolling out now to Chrome desktop users.

So Google is making Chrome itself an AI application platform.

That's the play. Keep users inside Google's ecosystem, make Gemini the default AI interaction layer for the web. The Hacker News discussion was skeptical, mainly about data collection motives and whether this further undercuts content creators by keeping users from interacting with websites directly. Valid concerns. But from a product strategy standpoint, embedding AI at the browser layer is smart. It meets users where they already are.

Quick hit on a potentially huge research development. A paper on Introspective Diffusion Language Models is generating serious excitement. Marcus, explain why this matters.

Current AI models generate text one token at a time, sequentially. Diffusion models can theoretically generate multiple tokens simultaneously, meaning much faster output. But diffusion language models have always been worse quality than autoregressive models. This paper cracks that problem.

How?

They identified the root cause as introspective consistency. Autoregressive models agree with their own generations, diffusion models don't. Their technique lets the model verify previous tokens while generating new ones in the same forward pass. They converted a Qwen model into a diffuser and got roughly three times higher throughput while matching quality. The benchmarks are strong, nearly seventy on AIME math, beating comparable diffusion models by over twenty-six points.

If this scales, what changes?

Everything about inference economics. Dramatically faster responses, lower serving costs. And because they converted an existing model rather than training from scratch, this could be adopted relatively quickly. It's early research, but if it holds up, it could fundamentally reshape how AI companies serve their models. Two hundred and fifty-seven points on Hacker News, and the technical community is genuinely excited.

One more quick hit. Fiverr left customer files, including complete tax returns, publicly accessible and indexed by Google. Five hundred and nine points on Hacker News.

Seven hours after disclosure, documents were still accessible via simple Google searches. Ten hours later, still live. The security report went through proper channels but the company was painfully slow to respond. When freelance platforms handle tax returns and legal documents and can't secure basic file permissions, that's a fundamental failure. Someone called it potentially business-ending, and they're not wrong.

Wednesday big picture. Marcus, OpenAI's investors are questioning an eight-hundred-billion-dollar bet, cybersecurity is becoming an AI arms race, vibe coding is creating real-world harm, and research papers are promising to make AI three times faster. What connects all of this?

Consequences are arriving. For two years, AI has been mostly about potential. Potential revenue, potential risks, potential breakthroughs. Now we're seeing actual outcomes. Investors want actual returns, not projections. Vulnerabilities need actual patching, not theoretical defenses. Vibe-coded apps are causing actual data breaches, not hypothetical ones. The industry is transitioning from what could happen to what is happening.

And the companies that prepared for this phase are pulling ahead.

Anthropic's revenue trajectory, Google's integration strategy, the efficiency revolution in smaller models. The winners are the ones who built for reality, not just for demos. The diffusion language model paper is exciting precisely because it solves a real economic problem, not just a benchmark number. We're past the era of impressive demos. Now it's about what works, what's safe, and what pays for itself.

Reality has entered the chat.

That's your AI in 15 for Wednesday, April 15, 2026. See you tomorrow.