# Ambient Advantage — April 24, 2026 *Friday · April 24, 2026 · [Episode page](https://podcast.ambient-advantage.ai/episodes/2026-04-24.html) · [Audio](https://storage.googleapis.com/ambient-advantage-podcast/2026-04-24-ambient-advantage.mp3)* [AVA] An AI model so dangerous it was restricted to forty organizations just got breached through a contractor's password and some URL guessing. That's the state of AI security in 2026. [JON] Yeah, we're going to unpack that one. Welcome to Ambient Advantage — I'm Jon, and this is Ava. It's Friday, April 24, 2026, and here's what matters in AI today. [AVA] And there is a lot that matters today, Jon. We've got OpenAI's new model positioned as an agent runtime, DeepSeek dropping frontier-adjacent performance at a tenth of US prices, Adobe killing its entire Experience Cloud brand, and a worker anxiety crisis that's becoming an actual business problem. [JON] But we're leading with that Anthropic security story because I think it's the one that should keep CISOs up tonight. So Ava, set the scene. What is Claude Mythos, and what happened? [AVA] Okay, so Anthropic has this model called Claude Mythos Preview. It's a cybersecurity-specific AI that can autonomously discover zero-day vulnerabilities across all major operating systems and browsers and then build multi-step exploits. These capabilities apparently emerged without explicit training, which is remarkable and terrifying in equal measure. Anthropic deemed it too dangerous for broad release and restricted access to about forty organizations under something called Project Glasswing — Apple, Amazon, Microsoft, Google, Cisco, major banks. [JON] So the crown jewels of AI cybersecurity, handed to a very small circle. And then what went wrong? [AVA] A Discord-based group of so-called "model hunters" gained access on the same day Mythos was publicly announced, April 7th. They exploited a third-party contractor's credentials and guessed URL conventions that were derived from a prior data leak at an AI training startup called Mercor. Not a sophisticated nation-state attack. Not a zero-day exploit. Credential reuse and URL guessing. [JON] That's... almost embarrassingly simple for the stakes involved. [AVA] It is, and that's exactly the point. Bloomberg got screenshots and a live demo from this group. They've had continuous access for weeks. Anthropic says they're investigating and found no evidence of core systems impact, but the damage to trust is done. And here's the business takeaway that I really want to hammer home. This isn't a story about Anthropic's internal security failing. Their own systems may be fine. This is the third-party vendor risk problem made visceral. [JON] So for enterprise risk leaders, what's the concrete action item? [AVA] Your supplier security assessments need to explicitly include AI model access controls. Full stop. If a contractor can access a model that can hack every major browser and OS, and that contractor's credentials are floating around from a previous breach at an unrelated company, your security perimeter is fiction. The attack surface for AI systems extends well beyond your own walls, and most enterprise risk frameworks haven't caught up to that reality yet. [JON] And worth noting, Microsoft just announced this week that it's integrating Mythos into its Security Development Lifecycle. So this model is becoming structural infrastructure for enterprise security, even as its own access controls just failed publicly. [AVA] Exactly. The question for security buyers is no longer "should we use AI for security" but "which frontier model's threat intelligence do we trust with our crown jewels, and how do we verify the entire chain of custody?" That's a fundamentally harder question than most organizations are equipped to answer today. [JON] Alright, let's move into the rundown. Ava, GPT-5.5 dropped yesterday. Give us the headline. [AVA] OpenAI launched GPT-5.5, internally nicknamed "Spud," and it's their clearest pivot from chat model to agent runtime. Greg Brockman says it can "figure out what needs to happen next" with minimal guidance. It scores nearly double Claude Opus 4.7 on post-doctoral math problems. But here's the catch — API pricing is double GPT-5.4 at five dollars per million input tokens and thirty dollars per million output. OpenAI says it's more token-efficient, but every enterprise running agentic workflows needs to revisit their ROI math immediately. [JON] And NVIDIA gave us the proof point on adoption, right? [AVA] Over ten thousand NVIDIA employees had early access across engineering, legal, marketing, finance, HR, and operations. NVIDIA's EVP called results "mind-blowing" and "life-changing." What matters here is the cross-functional deployment. This isn't just engineers using Codex anymore. When NVIDIA's finance and legal teams are on agentic AI, that's the competitive bar being set. Laggard enterprises are now structurally behind. [JON] Let's talk DeepSeek. V4 is out today and it sounds like a bombshell. [AVA] DeepSeek V4 is two models, both open-source under MIT license. V4-Pro has 1.6 trillion total parameters, trails GPT-5.4 by roughly three to six months on frontier benchmarks, but beats every open model in math, coding, and agentic tasks. And the pricing... Jon, V4-Pro is three dollars and forty-eight cents per million output tokens versus thirty dollars for GPT-5.5. V4-Flash is twenty-eight cents. That is a tenth the price for near-frontier performance. [JON] And it runs on Huawei chips, not Nvidia. [AVA] That's the geopolitical kicker. V4 runs on Huawei Ascend 950 chips, which is a milestone for China's compute independence. And the White House noticed. The OSTP issued a memo Thursday accusing Chinese firms of "industrial-scale" model distillation from US labs. They didn't name DeepSeek, but the timing speaks for itself. For enterprises evaluating open-source Chinese models, there's now a compliance risk dimension that didn't exist six months ago. [JON] Meanwhile DeepSeek is also raising its first external round? [AVA] Tencent and Alibaba are both circling. The valuation target doubled from ten billion to twenty billion plus in days. For context, this is a company owned by a hedge fund that generates minimal revenue because its models are free and open-source. But that investment will fund compute and infrastructure, which means V4's quality is only going to improve and the pricing pressure on US labs intensifies. [JON] Let's shift to Adobe. They killed Experience Cloud this week? [AVA] Killed it and replaced it with CX Enterprise, which is a full-stack agentic AI platform organized around a persistent "Coworker" agent that executes multi-step business goals autonomously with human oversight. It's built on open standards — MCP and Agent2Agent protocols — and interoperates with basically every major cloud and AI provider. Here's what caught my eye though. PwC is named explicitly as a system integrator leveraging Adobe's agentic capabilities for industry verticals. That's a direct revenue opportunity for our listeners in the consulting world. [JON] But there's a catch, right? [AVA] Adobe's own data says seventy-five percent of enterprises cite data integration as their number one AI blocker, and seventy-one percent cite talent gaps. So the platform decision isn't the hard part. The operating model and data maturity problem is what determines who actually benefits. That's where the real advisory work lives. [JON] And then there's Anthropic's valuation news, which is just staggering. [AVA] One trillion dollars implied on secondary markets. That's nearly triple their primary round from February, which was already three hundred and eighty billion. OpenAI, by contrast, trades at eight hundred and eighty billion on the same platform with a five-to-one seller-to-buyer ratio. The momentum swap is real. Anthropic's revenue went from nine billion annualized at end of 2025 to thirty to thirty-nine billion by March 2026. That's two hundred and thirty-three percent quarterly growth driven by Claude Code and enterprise API deals. There's speculative premium baked in, but there's real enterprise adoption underneath it. [JON] Alright, let's step back. The Bigger Picture. Ava, you've been looking at all of these stories together. What's the thread? [AVA] The thread is that we're watching the enterprise software stack get rewritten in real time, and the companies that understand this are acting while others are still evaluating. Look at what happened this week alone. OpenAI positioned its flagship model as an agent runtime, not a chat model. Adobe killed a billion-dollar brand to rebuild around autonomous agents. NVIDIA deployed agentic AI across every business function internally. OpenAI launched workspace agents that persist across enterprise tools. This isn't incremental improvement. This is a structural shift in how work gets done. [JON] And the human side of that shift? [AVA] That's the story people aren't talking about enough. Fewer than one in four workers feel their job is secure right now. There's a term for it — FOBO, Fear of Becoming Obsolete. And it's not just anxiety — companies are reporting measurable productivity declines tied to workforce distraction around AI displacement fears. Nearly eighty thousand tech layoffs in Q1 alone. [JON] Jensen Huang's line keeps echoing. "You won't lose your job to AI, you'll lose it to your coworker who uses it." [AVA] And that framing, while motivating for some, is deeply anxiety-inducing for most. Here's my take. FOBO isn't an HR problem. It's a change management problem that translates directly into failed transformation programs. Every agentic AI deployment needs to be paired with an explicit workforce transition narrative. The companies that do this will accelerate adoption. The companies that don't will see the implementation drag that undermines the ROI they're counting on. This is where consultants and business leaders earn their keep — not in the technology selection, but in bringing people along for the ride. [JON] Which brings it full circle to that Adobe data point. The blocker isn't the platform. It's the people and the data. [AVA] Every single time. The technology is now ahead of most organizations' ability to absorb it. That gap is the opportunity and the risk. [JON] What should people be watching next week? [AVA] Two things. First, watch for regulatory fallout from the OSTP memo on Chinese model distillation. If the US moves toward export controls on model weights or API access restrictions, that reshapes the open-source AI landscape overnight. Second, keep an eye on Anthropic's response to the Mythos breach. How they handle the disclosure and remediation will set expectations for how frontier AI companies manage security incidents going forward. It could become a template or a cautionary tale. [JON] I'll drop all the links and sources in the show notes for everything we covered today. [AVA] That's your Ambient Advantage for Friday, April 24, 2026. [JON] Share it with a colleague figuring out what AI means for their business. See you tomorrow.