Monday — May 25, 2026
Google just made every search bar an AI agent. Here is what that means.
Agent Security Audit
Act as a security architect. Review this enterprise AI agent workflow and identify three governance vulnerabilities related to data access, permission escalation, or unintended actions. Output each vulnerability with a one-sentence fix.
Google upgrades Search with Gemini 3.5 Flash AI Mode
Google quietly flipped a switch that makes every search query a potential AI interaction. Instead of links, users now get generated visuals, interactive charts, and even mini-apps built on the fly by Gemini 3.5 Flash. This is not a beta or an opt-in experiment. It is the default experience for billions of users. For builders, this means the classic SEO playbook is dead. Traffic used to flow to your site. Now it stays on Google. The new leverage is building tools that Google wants to embed, or creating content that feeds the model's reasoning chain. If you monetize through clicks, start diversifying into subscription or API-based revenue. If you build AI products, study how Google surfaces these mini-apps and find gaps where your tool can be the next embedded widget.
Flo's take: This is the biggest shift in search since Google launched. If you build on top of search traffic, your entire traffic model just changed overnight.
Anthropic hits profit, acquires OpenAI's SDK maker
Anthropic crossed into profitability faster than almost any AI infrastructure company. The $10.9B revenue number signals that enterprise contracts are real and sticky. But the acquisition of Stainless is the strategic move builders need to watch. Stainless built the SDK layer that OpenAI, Mistral, and others rely on for developer tooling. By owning that layer, Anthropic can shape how developers integrate any model, not just Claude. For builders, this means the SDK you use today might start favoring Anthropic's ecosystem tomorrow. The immediate action is to abstract your model calls behind a provider-agnostic interface. Do not hardcode any single SDK. Write your integration layer so you can swap out the underlying model provider in a day, not a month.
Flo's take: Anthropic just bought the pipes that connect everyone else's models. If you use any AI SDK, you are now indirectly on Anthropic's roadmap.
Microsoft launches Agent 365 governance plane
Microsoft's Agent 365 is not another model or chatbot. It is the infrastructure layer for controlling what agents can do, what data they access, and how they escalate permissions. At $15 per user per month, it is cheap enough that IT departments will buy it without a second thought. For builders, this creates a new standard. Enterprises will expect any third-party agent to plug into Agent 365's governance framework. If your agent cannot report its actions, log its decisions, or enforce role-based access controls, you will not get past procurement. The smart move is to build your agent with an open governance API that mirrors what Microsoft is doing. Support audit trails, permission scopes, and human-in-the-loop overrides out of the box. Make compliance a feature, not an afterthought.
Flo's take: Microsoft just put a price tag on AI agent safety. If you are selling agents to enterprises, you need to match this or get locked out.
Deep Dive
How to Build an Agent That Survives Enterprise Procurement
Enterprise procurement is the graveyard of AI startups. You can have the best model, the slickest UI, and a killer demo, but if your agent cannot pass a security review, you will never get a signed contract. The three stories above all point to the same reality: governance is the new moat. Google owns the front door, Anthropic owns the SDK layer, and Microsoft owns the control plane. You need to own your compliance story. Here is a step-by-step framework you can apply today. First, define your agent's permission model. Every action your agent takes should map to a specific scope: read data, write data, execute code, or call an external API. Use a simple enum, not free-form strings. Second, implement an audit trail. Every decision your agent makes must be logged with a timestamp, the user who authorized it, the input, the output, and the model version. Store this in an append-only log, not a mutable database. Third, add a human-in-the-loop override. For any action that costs money, deletes data, or sends a message to a customer, require explicit approval. Build this as a webhook that pauses execution until a human clicks confirm. Fourth, expose a compliance API. Microsoft's Agent 365 and similar tools will query your agent for its governance status. Give them a /health endpoint that returns your permission scopes, audit log integrity, and override status. Fifth, test with a red team. Before you pitch to any enterprise, run a simulated attack. Try to make your agent leak a prompt, access forbidden data, or execute an unauthorized command. Fix those gaps. Then document them. Procurement teams love seeing that you have already thought about the edge cases. The builders who treat governance as a feature, not a tax, will win the enterprise deals. The ones who ignore it will be stuck selling to solo developers on Stripe.
Go build something that scares you.