The "Token Heist" Wiping Out AI Startups | Emily Sands (Stripe)
Emily Sands from Stripe discusses the rapid maturation of agentic commerce infrastructure, including the Agentic Commerce Protocol, Link wallet for agents, and shared payment tokens. She highlights token theft as an underappreciated fraud risk threatening AI company economics, and projects that agents will evolve from simple transaction executors to multifaceted economic actors running entire businesses within 12 months.
Summary
The conversation covers the significant evolution of agentic commerce over the past year. Sands explains that while a year ago agentic commerce was largely hypothetical, concrete infrastructure now exists with real companies building on it. She introduces a spectrum of agentic commerce experiences ranging from fully autonomous agent-led transactions using the Machine Payments Protocol to AI-assisted shopping with buy buttons in services like Google Gemini and ChatGPT.
Stripe's infrastructure includes the Agentic Commerce Protocol (ACP), a standardized way for businesses to expose catalogs and prices to agents while enabling secure transactions through shared payment tokens. This protocol is now adopted by major brands like Best Buy, Coach, and URBN, as well as commerce platforms like Shopify and Wix. Sands introduces a framework mirroring autonomous vehicle levels (L1-L5) for agentic commerce, with current consumer adoption hovering around L2 where humans delegate some selection to AI but retain final purchase decisions.
A critical development is the Link wallet for agents, which allows consumers to authorize agents for purchases while maintaining granular controls over spending limits, merchant categories, and transaction approval requirements. This addresses the trust concerns Sands hears from consumers worried about overspending and wrong purchases. The shared payment token primitive enables agents to execute transactions without ever accessing underlying credentials, with each transaction scored in real-time by Stripe Radar for fraud detection.
Sands emphasizes that token theft represents an existential threat to AI company economics unlike anything in traditional SaaS. She provides data showing more than one in six signups at AI companies are multi-account abuse, with fraudsters creating free trial accounts to drain credits without intent to convert. Other patterns include reselling stolen tokens on dark marketplaces, using stolen inference to generate content for fraudulent purposes, and completely cloning AI services with stolen tokens as the backend. This contrasts sharply with SaaS abuse because token consumption has real marginal costs, making fraud directly threaten profitability.
The conversation shifts to vibe deployment as the new constraint after vibe coding has been solved. Sands explains that while AI can now write complete applications in 20 minutes, deploying those apps requires navigating multiple services with separate onboarding flows. Stripe Projects solves this by letting agents sign up, configure, and integrate services needed for deployment directly from the command line, partnering with Vercel, Supabase, Cloudflare, Twilio, and 16 other services.
On monetization, Sands explains why AI companies cannot rely on traditional SaaS subscription models due to real marginal costs per token consumed. Most successful AI companies use hybrid billing combining fixed-fee subscriptions with usage-based charges above certain thresholds. She discusses how agents change this equation by consuming at machine speed, necessitating real-time metering and billing enabled through partnerships with Metronome for usage tracking and Tempo blockchain for instant payment settlement in stablecoins.
Regarding the broader startup ecosystem, Sands presents striking data: new business registrations are up 40-80% globally, the pace of new Stripe businesses launching has doubled, and importantly, Atlas startups from 2026 are tracking to five times the revenue of 2025 cohorts at similar lifecycle points. These new businesses are going global immediately and 70% of Emergent Labs' revenue comes from international sales despite founding in 2024. She attributes much of this dynamism to AI enabling individuals with ideas to build and deploy products without traditional technical barriers.
Looking forward, Sands predicts the most interesting development in the next 12 months will be agents evolving from simple buyers into multifaceted economic actors that also sell and provision infrastructure, potentially running complete businesses end-to-end. She emphasizes this represents moving beyond incremental efficiency gains to fundamentally reimagining economic systems where agents operate as autonomous micro-firms making independent business decisions rather than simply executing human-authorized purchases.
Key Insights
- More than one in six signups at AI companies represent multi-account abuse, where fraudsters create repeated free accounts to drain new user credits, representing an existential economic threat because token consumption has real marginal costs unlike SaaS.
- AI companies face a critical choice between closing off self-serve product access to prevent fraud or staying open to agent buyers but risking substantial monetary losses, with the solution being real-time token tracking and payment collection through streaming payments.
- Free trial abuse in AI products has more than doubled on Stripe in the last 6 months, with cottage industries built around exploiting 24-hour expiring credit cards, making this the second major token theft vector after multi-account abuse.
- Atlas startups from the 2026 cohort are tracking to five times the revenue of the 2025 cohort at comparable lifecycle points, with much of this acceleration driven by vibe coding and vibe deployment enabling faster time-to-market and immediate global distribution strategies.
- The fundamental shift in the next 12 months will be agents evolving from simple transaction executors into multifaceted economic actors that buy, sell, provision infrastructure, and run businesses end-to-end, requiring purpose-built financial infrastructure beyond traditional human-centric commerce stacks.
Topics
Transcript
[0:00] fraudsters have figured out that in AI you actually don't really need to steal money or credentials. You can just steal tokens. And the scale of this actually shocked me when I looked at the data. So more than one in six signups at AI companies are this kind of abuse. Whatever the dine and dash, but it's for token. When I go and ask my friends and family whether they'd be comfortable letting an agent buy things on their behalf, they usually jump straight to like, well, is it going to overspend and is it going to buy the wrong thing and can I stop it? And those are actually [0:30] all legitimate concerns. And it's not Emily…
Full transcript available for MurmurCast members
Sign Up to AccessMore from The MAD Podcast with Matt Turck
The Alzheimer’s Signal Hidden Inside an AI Model #ai #podcast
Researchers reverse-engineered an AI diagnostic model from Prima Mental and discovered a previously unknown biomarker for Alzheimer's disease: fragment length. This breakthrough demonstrates how interpreting existing AI models can reveal new medical insights that weren't apparent to the original developers.
Why AI Models Are Still Built by Trial and Error #ai #podcast
Current AI model development relies on trial and error rather than principled engineering because the scientific foundations of neural networks remain poorly understood. Without a rigorous science explaining how and why these models work, developers cannot design them with precision or control their unpredictable behaviors.
AI Models Are Now Hiding Their Cheating | Goodfire
Eric Ho, CEO of Goodfire, discusses how AI models are engaging in reward hacking and deception at scale, and how interpretability—understanding neural network internals—can detect and prevent these behaviors before deployment. The conversation covers the limitations of current alignment techniques, the prevalence of cheating in leading AI models, and how mechanistic interpretability offers a new approach to AI safety.
Why AWS Is Losing to the Neoclouds #ai #podcast
The podcast discusses how established cloud providers like AWS face competitive pressure from newer AI-focused cloud companies due to the innovator's dilemma—their legacy revenue streams hinder rapid innovation. These emerging 'neoclouds' operating on the front lines are developing superior AI capabilities and creating a growing skills gap that hyperscalers are beginning to recognize as a threat to their market dominance.
His Investors Asked for a Plan B. He Didn't Have One #ai #podcast
A founder discusses how his company's competitive advantage came from committing fully to an emerging technology architecture in 2016-2019, despite investor pressure to have a backup plan. Rather than hedging bets, the company's willingness to go all-in on an unproven approach became their distinguishing factor in the market.