MrBeast bought a bank. Not quite. He acquired Step, a teen-focused fintech with 7 million customers, built on the rent-a-charter model via Evolve Bank & Trust. But the real story isn't a fintech acquisition. It's the Feastables playbook applied to finance: take a commoditized product, wrap it in distribution nobody else can match, and own the relationship.

MrBeast's core competency is attention science. Starting in 2012, he reverse-engineered the YouTube algorithm to build 1.4 billion subscribers. That's not a follower count; it's a marketing funnel that most banks couldn't replicate with billions in ad spend. Feastables proved the model. A commoditized product (chocolate bars) wrapped in unmatched distribution (460M subscribers), with obsessive product quality driving repeat purchases. As he puts it: "Virality gets the first purchase, quality gets the second." Feastables generated $250M in revenue in 2024 with distribution across Walmart, Target, and 7-Eleven.

What makes Step interesting is its Secured Card program. To the user, it looks and acts like a debit card. Technically, it's a secured credit card. Every transaction gets reported to credit bureaus as a paid-on-time credit payment. Teens enter college with a 700+ credit score already built. That credit score is the stickiness mechanism. A chocolate bar is a one-off transaction at Walmart. A bank account becomes the customer's financial identity. The longer you stay, the more valuable your credit history becomes, and the harder it is to leave.

The core of the attention arbitrage is that the form of trust is different. Traditional banks ask for trust at the point of sale. They need to convince through ads, brand reputation, and marble lobbies. Creator trust arrives pre-loaded, built through years of content and parasocial relationships. That's why the CAC math is fundamentally different. In 2026, neobank customer acquisition costs have soared past $150-300 due to ad-market saturation. MrBeast likely displaces 90%+ of ad spend singlehandedly.

The risks are real, though. Step runs on Evolve Bank & Trust, which was at the center of the Synapse bankruptcy and is under a Federal Reserve Consent Order. Competitors like SoFi, Varo, and Monzo have secured their own banking charters, moving past the BaaS model. And a brand known for "I Blew Up My House" managing teenagers' savings means a single technical glitch could spiral into a PR crisis that no apology video can fix.

Still, the core question stands. Most people stay with their first bank for decades. Capture a 13-year-old user today, and by the time they need mortgages and investment accounts, Step owns that relationship. The threat to banks isn't that MrBeast is a better banker. It's that he's a better onramp.

BCG published Commerce Everywhere, Agents in Charge. The central diagnosis is that two structural shifts are happening simultaneously. First, commerce is moving from channels to surfaces. Second, AI agents are emerging as the organizing layer for buying decisions.

Commerce escaping channels means purchasing no longer happens at a destination like a website or app. Social feeds become storefronts, media content leads directly to checkout, and AR experiences resolve pre-purchase uncertainty. TikTok Shop's GMV has more than doubled year-over-year and is forecast to surpass $30 billion by 2028. Retail media ad spending is on track to exceed $100 billion globally. The global e-commerce market itself is projected to grow from roughly $25 trillion in 2024 to over $80 trillion by 2030.

More surfaces means more complexity. Managing dozens of touchpoints, offers, and decisions becomes impractical for consumers. Enter AI agents. Consumers state intent ("find me a quiet dishwasher," "plan a long weekend trip") and the agent handles research, comparison, optimization, basket assembly, and checkout. BCG calls this a "seismic shift" in how commerce is conducted.

BCG's proposed architecture has three layers:

• Surfaces become inputs. Feeds, media, and AR experiences generate purchase intent signals without requiring checkout.

• AI agents orchestrate action. They shop across brand sites, marketplaces, social shops, and local inventory to fulfill intent.

• Universal carts enable execution. A single infrastructure layer where agents assemble, optimize, and transact baskets across multiple merchants.

The universal cart layer is interesting. Shopify's Commerce for Agents toolkit, launched in 2025, enables AI assistants to search large product catalogs, assemble multi-seller baskets, and complete transactions within conversational experiences. Platforms like Firmly.ai extend this further with multi-merchant, multi-cart checkout. Whoever operates this layer gains visibility into how baskets form across retailers, where substitutions occur, and how consumers trade off speed, price, and convenience.

I've covered agentic commerce in a previous weekly review, where the analysis pointed to infrastructure layers like Shopify and Stripe as primary beneficiaries. BCG's piece goes a step further, asking what strategy brands should adopt in an agent-mediated market. The "destination game" (drive customers to owned channels) or the "evaluation game" (get recommended by agents). BCG argues most brands will choose the evaluation game. Agents bypass offers that are ambiguous, incomplete, or hard to execute. What gets selected is not what appeals to humans, but what machines can easily understand.

The strategic question shifts from "Which channel are we optimizing?" to "When AI agents initiate and optimize buying decisions across every surface, how does our brand get chosen?"

Paradigm and OpenAI jointly released EVMbench, a benchmark assessing whether AI agents can detect and exploit smart contract vulnerabilities within simulated blockchain environments. While not technically an RL (reinforcement learning) environment, its structure parallels RL systems, meaning it could facilitate future model training.

Here's how it works. Real vulnerabilities sourced from Code4rena auditing competitions are deployed on forked blockchain instances. AI models are given three tasks: detect, patch, and exploit. Eight models were tested. The most striking finding is the asymmetry. Models performed significantly better at exploiting vulnerabilities than at detecting or patching them.

This asymmetry is explained by differences in reward structure. Exploitation is binary: you either stole the funds or you didn't. Clear success/failure signals make it easy for models to optimize. Detection, by contrast, requires exhaustive search through codebases. Patching is a compound task that demands both finding and fixing vulnerabilities.

Why does this matter? Roughly $100 billion currently sits on-chain, and the total crypto asset class is valued at $2.3 trillion. Treasury Secretary Scott Bessent has stated the goal of putting $3 trillion in stablecoins on-chain by 2030. In this context, the fact that frontier models can autonomously discover and exploit smart contract vulnerabilities end-to-end using publicly accessible APIs is not merely an academic finding.

Zooming out, EVMbench is part of a broader trend around RL environments. These are simulated software sandboxes that improve post-training model capabilities. Through verifiable, step-by-step tasks, they reduce reliance on human evaluation while building agentic competence. Anthropic is reportedly investing roughly $1 billion in RL environment acquisition. Companies like Hud are building tools that convert any software into RL environments. Prime Intellect is pursuing decentralized AGI development through an open-source RL environment hub.

EVMbench's evaluation harness and dataset have been fully open-sourced. This can accelerate defensive security research, but it also hands the same tools to malicious actors. The concerns mirror those raised by Anthropic's Red Team in biorisk research, with one critical difference: in decentralized finance, the time gap between discovering a vulnerability and draining funds is near-zero.

As more value moves on-chain, smart contract security becomes a far more urgent infrastructure challenge than incremental benchmark improvements. The asymmetry where models are better at attacking than defending means defensive tooling and audit processes must keep pace with the rate at which offensive capabilities advance.

MrBeast bought a bank. Not quite. He acquired Step, a teen-focused fintech with 7 million customers, built on the rent-a-charter model via Evolve Bank & Trust. But the real story isn't a fintech acquisition. It's the Feastables playbook applied to finance: take a commoditized product, wrap it in distribution nobody else can match, and own the relationship.

MrBeast's core competency is attention science. Starting in 2012, he reverse-engineered the YouTube algorithm to build 1.4 billion subscribers. That's not a follower count; it's a marketing funnel that most banks couldn't replicate with billions in ad spend. Feastables proved the model. A commoditized product (chocolate bars) wrapped in unmatched distribution (460M subscribers), with obsessive product quality driving repeat purchases. As he puts it: "Virality gets the first purchase, quality gets the second." Feastables generated $250M in revenue in 2024 with distribution across Walmart, Target, and 7-Eleven.

What makes Step interesting is its Secured Card program. To the user, it looks and acts like a debit card. Technically, it's a secured credit card. Every transaction gets reported to credit bureaus as a paid-on-time credit payment. Teens enter college with a 700+ credit score already built. That credit score is the stickiness mechanism. A chocolate bar is a one-off transaction at Walmart. A bank account becomes the customer's financial identity. The longer you stay, the more valuable your credit history becomes, and the harder it is to leave.

The core of the attention arbitrage is that the form of trust is different. Traditional banks ask for trust at the point of sale. They need to convince through ads, brand reputation, and marble lobbies. Creator trust arrives pre-loaded, built through years of content and parasocial relationships. That's why the CAC math is fundamentally different. In 2026, neobank customer acquisition costs have soared past $150-300 due to ad-market saturation. MrBeast likely displaces 90%+ of ad spend singlehandedly.

The risks are real, though. Step runs on Evolve Bank & Trust, which was at the center of the Synapse bankruptcy and is under a Federal Reserve Consent Order. Competitors like SoFi, Varo, and Monzo have secured their own banking charters, moving past the BaaS model. And a brand known for "I Blew Up My House" managing teenagers' savings means a single technical glitch could spiral into a PR crisis that no apology video can fix.

Still, the core question stands. Most people stay with their first bank for decades. Capture a 13-year-old user today, and by the time they need mortgages and investment accounts, Step owns that relationship. The threat to banks isn't that MrBeast is a better banker. It's that he's a better onramp.

BCG published Commerce Everywhere, Agents in Charge. The central diagnosis is that two structural shifts are happening simultaneously. First, commerce is moving from channels to surfaces. Second, AI agents are emerging as the organizing layer for buying decisions.

Commerce escaping channels means purchasing no longer happens at a destination like a website or app. Social feeds become storefronts, media content leads directly to checkout, and AR experiences resolve pre-purchase uncertainty. TikTok Shop's GMV has more than doubled year-over-year and is forecast to surpass $30 billion by 2028. Retail media ad spending is on track to exceed $100 billion globally. The global e-commerce market itself is projected to grow from roughly $25 trillion in 2024 to over $80 trillion by 2030.

More surfaces means more complexity. Managing dozens of touchpoints, offers, and decisions becomes impractical for consumers. Enter AI agents. Consumers state intent ("find me a quiet dishwasher," "plan a long weekend trip") and the agent handles research, comparison, optimization, basket assembly, and checkout. BCG calls this a "seismic shift" in how commerce is conducted.

BCG's proposed architecture has three layers:

• Surfaces become inputs. Feeds, media, and AR experiences generate purchase intent signals without requiring checkout.

• AI agents orchestrate action. They shop across brand sites, marketplaces, social shops, and local inventory to fulfill intent.

• Universal carts enable execution. A single infrastructure layer where agents assemble, optimize, and transact baskets across multiple merchants.

The universal cart layer is interesting. Shopify's Commerce for Agents toolkit, launched in 2025, enables AI assistants to search large product catalogs, assemble multi-seller baskets, and complete transactions within conversational experiences. Platforms like Firmly.ai extend this further with multi-merchant, multi-cart checkout. Whoever operates this layer gains visibility into how baskets form across retailers, where substitutions occur, and how consumers trade off speed, price, and convenience.

I've covered agentic commerce in a previous weekly review, where the analysis pointed to infrastructure layers like Shopify and Stripe as primary beneficiaries. BCG's piece goes a step further, asking what strategy brands should adopt in an agent-mediated market. The "destination game" (drive customers to owned channels) or the "evaluation game" (get recommended by agents). BCG argues most brands will choose the evaluation game. Agents bypass offers that are ambiguous, incomplete, or hard to execute. What gets selected is not what appeals to humans, but what machines can easily understand.

The strategic question shifts from "Which channel are we optimizing?" to "When AI agents initiate and optimize buying decisions across every surface, how does our brand get chosen?"

Paradigm and OpenAI jointly released EVMbench, a benchmark assessing whether AI agents can detect and exploit smart contract vulnerabilities within simulated blockchain environments. While not technically an RL (reinforcement learning) environment, its structure parallels RL systems, meaning it could facilitate future model training.

Here's how it works. Real vulnerabilities sourced from Code4rena auditing competitions are deployed on forked blockchain instances. AI models are given three tasks: detect, patch, and exploit. Eight models were tested. The most striking finding is the asymmetry. Models performed significantly better at exploiting vulnerabilities than at detecting or patching them.

This asymmetry is explained by differences in reward structure. Exploitation is binary: you either stole the funds or you didn't. Clear success/failure signals make it easy for models to optimize. Detection, by contrast, requires exhaustive search through codebases. Patching is a compound task that demands both finding and fixing vulnerabilities.

Why does this matter? Roughly $100 billion currently sits on-chain, and the total crypto asset class is valued at $2.3 trillion. Treasury Secretary Scott Bessent has stated the goal of putting $3 trillion in stablecoins on-chain by 2030. In this context, the fact that frontier models can autonomously discover and exploit smart contract vulnerabilities end-to-end using publicly accessible APIs is not merely an academic finding.

Zooming out, EVMbench is part of a broader trend around RL environments. These are simulated software sandboxes that improve post-training model capabilities. Through verifiable, step-by-step tasks, they reduce reliance on human evaluation while building agentic competence. Anthropic is reportedly investing roughly $1 billion in RL environment acquisition. Companies like Hud are building tools that convert any software into RL environments. Prime Intellect is pursuing decentralized AGI development through an open-source RL environment hub.

EVMbench's evaluation harness and dataset have been fully open-sourced. This can accelerate defensive security research, but it also hands the same tools to malicious actors. The concerns mirror those raised by Anthropic's Red Team in biorisk research, with one critical difference: in decentralized finance, the time gap between discovering a vulnerability and draining funds is near-zero.

As more value moves on-chain, smart contract security becomes a far more urgent infrastructure challenge than incremental benchmark improvements. The asymmetry where models are better at attacking than defending means defensive tooling and audit processes must keep pace with the rate at which offensive capabilities advance.