# Ambient Advantage — May 29, 2026 *Friday · May 29, 2026 · [Episode page](https://podcast.ambient-advantage.ai/episodes/2026-05-29.html) · [Audio](https://storage.googleapis.com/ambient-advantage-podcast/2026-05-29-ambient-advantage.mp3)* [AVA] Robinhood just gave AI agents their own wallets, their own credit cards, and the ability to trade stocks on behalf of 27 million customers. The age of AI as assistant is officially over. [JON] Yeah, we need to talk about that one. [JON] Welcome to Ambient Advantage — I'm Jon, and this is Ava. It's Friday, May 29, 2026, and here's what matters in AI today. We've got a packed show. Robinhood letting AI agents trade autonomously, Cognition's Devin hitting a billion-dollar raise, Claude Opus 4.8 dropping, ChatGPT running ads, the FBI creating a whole new crime category for AI fraud, and Ethan Mollick telling a room full of corporate leaders that nobody has a playbook. Let's get into it. [AVA] Let's start with Robinhood, because this is one of those stories that sounds like a headline from 2030 but it happened on Tuesday. [JON] Walk us through it. [AVA] Robinhood launched what they're calling Agentic Trading in beta. Their 27 million funded customers can now connect AI agents — from Claude, from ChatGPT — to sandboxed trading accounts. The agents can analyze concentration risk, execute trades, dig through analyst notes to find opportunities. And they do this through Robinhood's MCP service, so there's a structured protocol governing how the agent interacts with the platform. [JON] Okay, that's already a lot. But you mentioned a credit card? [AVA] Right. Robinhood also launched an Agentic Credit Card — a virtual Robinhood Gold Card with 3% cash back and configurable spending limits specifically designed for AI agents. So your agent doesn't just advise you on what to buy. It has a wallet. It has a card. It can spend. [JON] So we've gone from "AI helps me think about my portfolio" to "AI acts on my behalf with real money." That's a fundamentally different thing. [AVA] It is, and that's exactly why this is our lead story today. The shift from assist to act is the defining transition in agentic AI right now, and Robinhood just made it concrete in one of the most regulated industries on Earth. They've built in fraud detection, sandboxed accounts, spending limits. But here's what enterprise leaders should be paying attention to: the governance architecture. Permissions, spending caps, audit logs, rollback buttons, and yes ... a panic switch. [JON] A literal panic switch? [AVA] Effectively, yes. The ability to freeze agent activity instantly. And that's not a nice-to-have feature. That is the minimum viable compliance layer for any agentic system touching real assets. If you're building agent workflows inside your enterprise that touch procurement, finance, customer data — you need the same architecture. Robinhood is the canary here. [JON] So the message for enterprise buyers isn't just "look what Robinhood did." It's "this is what your internal agent governance needs to look like." [AVA] Exactly. Crypto support is coming next, options are planned. This is only going to accelerate. I'll drop a great technical breakdown of how the MCP-connected wallets actually work in the show notes. [JON] Let's move into the rundown. Ava, Cognition raised over a billion dollars this week. Give us the number and then give us the story behind the number. [AVA] Cognition, the company behind the AI software engineer Devin, raised over a billion dollars at a 26-billion-dollar post-money valuation. But the real story is the revenue trajectory. They went from 37 million to 492 million in twelve months. That's a 13x increase. And here's the stat that stopped me: 89% of all code committed at Cognition is now committed by Devin. In December 2025, that number was 13%. [JON] 89% of their own code is written by their own AI. [AVA] Yes. And their enterprise customer list tells you this isn't a toy. Goldman Sachs, Mercedes-Benz, Dell, Santander, the US Army, the US Navy. Mercedes reportedly reduced an eight-month legacy modernization project to eight days. Brazilian bank Itaú is fixing 70% of security vulnerabilities automatically with Devin. If your enterprise is still evaluating whether agentic coding belongs in your engineering workflow, you are running out of time to form an opinion. [JON] Next up, Anthropic shipped Claude Opus 4.8. It feels like we just talked about 4.7. [AVA] 41 days ago, to be exact. Opus 4.8 scores 88.6% on SWE-bench Verified and sits 121 Elo points ahead of GPT-5.5 on reasoning benchmarks. The big new capability is dynamic workflows — the model can plan a task and spin up hundreds of parallel sub-agents in a single session. Claude Code with Opus 4.8 can handle codebase-wide migrations across hundreds of thousands of lines, from planning all the way to merge. [JON] What about pricing? [AVA] Flat. Standard pricing unchanged. And the new Fast Mode, which runs Opus 4.8 at 2.5x speed, actually costs a third of what it did for earlier models. Ten dollars per million input tokens, fifty for output. But here's the detail I want enterprise buyers to notice: Anthropic says Opus 4.8 is four times less likely than 4.7 to let flaws in its own code pass without flagging them. If you're the only human reviewing AI-generated code before it ships to production, that is the most valuable improvement in this release. [JON] Alright, staying in the OpenAI universe for a second — ChatGPT is now running ads? [AVA] It is. OpenAI launched an advertising system inside ChatGPT in mid-April. The initial rollout targets free and Go tier users in Australia, New Zealand, and Canada. And according to internal documents from their ad partner StackAdapt, the targeting mechanism is based on what they call "prompt relevance." Meaning ads are served based on the content of your prompts. [JON] So every conversation you have with ChatGPT could be used to target you with ads. [AVA] On the free tier, yes. And Anthropic has publicly committed to keeping Claude ad-free permanently. This is now a structural business model split between the two leading platforms. But here's the enterprise angle that matters right now: if your employees are using free ChatGPT accounts with company data, those work prompts are now ad-targeting signals. The shadow IT risk just got materially higher. Enterprise licensing conversations should now explicitly include data-use policies, not just output quality. [JON] That's a great point. Let's talk security, because we've got two alarming stories this week. The FBI and the jailbroken Gemini hack. [AVA] Two stories that belong together. First, the FBI has formally designated AI fraud as its own crime category for the first time in the IC3's 26-year history. They logged 22,364 complaints costing Americans nearly 893 million dollars, and the FBI itself says that's likely an undercount. Interpol reported that AI-aided financial fraud schemes are 4.5 times more profitable than those without AI. [JON] And the Gemini incident? [AVA] A Russian-speaking threat actor used a jailbroken version of Google's Gemini to steal admin credentials in a documented attack. Separately, researchers found that Claude leaked AWS keys in 24 of 25 test cases when adversarially prompted. Both incidents land as MCP adoption is accelerating and agents are getting more tool access. This is a two-alarm fire. Agentic systems with access to cloud credentials, databases, or APIs must be treated as a distinct threat surface. Zero-trust principles apply to AI agents just as they do to human developers. Least privilege, secrets managers, no hardcoded keys in agent context. Period. [JON] Last one in the rundown — OpenAI launched Secure MCP Tunnels. This sounds technical, but Ava, I think this might be one of the most consequential things we cover today. [AVA] It might be. The single biggest blocker to enterprise MCP adoption has been the security objection: "We can't expose internal systems to the public internet." OpenAI just removed that objection. Secure MCP Tunnel lets you connect private MCP servers to ChatGPT, Codex, and their Responses API without opening inbound firewall ports. A tunnel client runs inside your network, opens an outbound HTTPS connection to OpenAI, and pulls queued MCP work. It supports outbound proxies, custom CA bundles, client certificates, mTLS on the MCP side. Greg Brockman called it "bring your own MCP servers." For regulated industries, this is the credible path to connecting internal databases, ERP systems, and knowledge bases to AI agents without a security review nightmare. [JON] Let's zoom out to the bigger picture. What ties all of this together? [AVA] This week's stories sketch the outlines of a new enterprise reality, and the word that keeps coming up is "principal." Agents are no longer prototypes. They are principals. They're trading stocks. They're writing 89% of production code. They're connecting to internal systems through MCP tunnels that bypass the firewall objection entirely. And the governance architecture most enterprises have built for AI ... it was designed for tools that assist humans. Content policies, acceptable-use guidelines, output review checklists. That governance is almost entirely unfit for tools that act on behalf of humans. [JON] So what does fit? [AVA] Robinhood's panic switch, Cognition's audit logs, OpenAI's MCP tunnel with mTLS — these aren't features. They are the minimum viable compliance layer for agentic AI touching real assets. And I want to bring in Ethan Mollick here, because he said something this week that I think every executive needs to hear. He told a room of corporate leaders at the New York Public Library: "Nobody knows anything. I talk to AI labs, famous people, CEOs all the time, and nobody knows anything. We're all making this up as we go along. Anyone who says they have the playbook — they're lying to you." [JON] That's from the most-read academic voice in applied AI. [AVA] It is. And he's right. The honest enterprise posture right now is iterative experimentation, not strategic lock-in. If your board is demanding a definitive five-year AI strategy by Q3, they may be asking for something no one on Earth can legitimately deliver. The companies that will win are the ones experimenting fast, governing well, and staying humble about what they don't know. [JON] And Sam Altman seems to have gotten that memo too — walking back his AI jobs apocalypse warnings this week. [AVA] Conveniently timed, given OpenAI is reportedly weeks from filing for an IPO targeting a trillion-dollar valuation. Meanwhile, tech layoffs through May 2026 have passed 115,000 — already approaching the full-year 2025 total — with Meta, Amazon, and Snap citing AI as a driver. The job-replacement panic is overstated in the short term, but the restructuring happening at Goldman, HSBC, and Cisco right now is very real. Don't confuse Altman's optimism tour with an all-clear signal. [JON] What should people be watching next week? [AVA] Two things. First, watch how quickly other financial platforms follow Robinhood's agentic trading model. If a major bank or brokerage announces something similar in the next thirty days, we're in a full-blown agentic finance race. Second, keep an eye on the Biohub protein world model. They released open models this week that can design protein binders validated in actual lab experiments. If you're in pharma or biotech, your R&D capital allocation assumptions may need revisiting. I'll drop the Biohub announcement link in the show notes. [AVA] That's your Ambient Advantage for Friday, May 29, 2026. [JON] Share it with a colleague figuring out what AI means for their business. See you tomorrow.