
Live Builders of Charlotte recording at the Builder Block Party
A lot happened in cyber last week, and most of it happened in about 5 days. On Tuesday the 20th, 8 security certificates that browsers use to trust websites turned out to be compromised, because the keys behind them could be worked out from the certificates themselves. On Wednesday, PHP (the language behind about 70% of the world's websites) rushed out patches for two critical flaws that attackers were already trying to exploit, and Microsoft published a report on a China-aligned group using an open-source AI hacking framework against more than 2,250 organizations. By Thursday, banks were telling reporters they'd found customers' personal AI agents poking around inside their accounts.
Azeem Azhar (who writes Exponential View) called it "the internet's Wuhan moment." His point was that it feels like January 2020 again.. something is spreading fast, a few people are worried and most people haven't noticed yet.
On Saturday we recorded the first live episode of Builders of Charlotte at the Builder Block Party at the CLT Startup House, with my friends Charles D'Andrea, Colin Thornton and Alan Corvalis. Early on I asked the room who'd been dealing with the cyber stuff from the last couple of days. Nobody raised a hand, and that scared me just a little bit. The whole episode is up now if you want to watch it.
1. Attackers got AI agents before most companies did
In the Microsoft report, the group used a framework that can plug in any AI model and run the boring parts of an attack on its own, then hit more than 2,250 organizations at 2,900 sign-in attempts a minute. In multiple cases it went from first contact to admin-level access in under 3 hours. Microsoft says most of the attempts failed, which is good news, but some got in.
The other half of the story is how fast the holes are being found. Azeem points out that one researcher found the PHP flaws in about a week with AI helping, work that would've taken a world-class team months not long ago. So the cost of finding a way in has dropped a ton, and it's dropped for everyone, including people who aren't on your side.
Colin pointed out on the panel that most attackers aren't coming after you personally. They're going after platforms and apps, because that's where the data is. If your company runs on a website, a customer portal or an app someone vibe coded over a weekend, you're on that list whether you like it or not.
The way I think about it is leaving something valuable on the front seat of your car with the window rolled down. You'll never make the car impossible to break into, BUT you can make it a lot less likely someone picks yours.

2. Your AI agents are part of the attack surface now
One of the builders in the audience on Saturday volunteers for a nonprofit where partnership emails were sitting unread (some of them literally offering money), so they started building an agent to answer them. While they were putting it together, they did what we'd just been talking about and asked AI whether it had any vulnerabilities. The answer was that anyone who forwarded an email the agent was copied on could ask it to pull the nonprofit's filings out of Google Drive. They immediately took it down to avoid data getting leaked.
That's prompt injection in plain terms. An agent that reads email, documents or web pages will also read instructions hidden inside them, and it can't always tell your instructions from someone else's. If that same agent also holds the keys to your files, you've built a way in for whoever writes the most convincing email.

The banks story is the same problem from the other direction. Customers are handing personal AI agents their bank logins, and the banks are finding those agents doing things like calling in a complaint without checking with the person who owns the account. Nobody was draining accounts, but the White House's AI advisor (David Sacks) suggested banks offer read-only access as the middle ground, and that full write access "should be approached carefully."
Charles's take on the panel was to think about an agent the way you'd think about a new employee, and ask what the worst thing a disgruntled employee could do in that role would be, how fast you'd notice and how you'd recover. That question works a lot better for agents than thinking about them like another piece of software you bought.
3. The trust checklist in the playbook doubles as a security checklist
Last week I shared 4 pages from the 2027 AI Strategy Playbook (the free 55-page playbook I wrote for leadership teams planning AI into next year's budget), including 3 of the 8 boxes from the trust checklist in chapter 08. It was written to keep AI output accurate enough that people would trust it, and reading it again after last week, almost every box is also a security control.

4 of them, read with agents in mind:
Access starts scoped and read-only, with no delete. Turning on write access is a decision someone signs. That's exactly what the White House suggested to the banks, and it would've made the nonprofit agent a lot less dangerous.
Onboard it like a new hire. Write down where the data is, what its job is, which tools it can use and what good looks like. If you can't write down what it's allowed to touch, keep it away from everything until you can.
AI drafts, a person approves. Nothing reaches a customer (or a bank, or a vendor) until the workflow has earned it, and a forwarded email asking for your filings is exactly what that approval step catches.
When it doesn't know, it says so and hands off to a named person. An agent that's allowed to say "I'm not sure, let me get someone" is a lot harder to talk into something.
One rule I'd add from our own work came out of designing a monthly security audit across our projects at StealthX. The agent that reads untrusted content (email, a client's repo, a web page) shouldn't be the one holding the passwords and keys, so in our design the part that holds credentials doesn't use AI at all and the AI part never sees a credential. Even Anthropic's own code review tool says in its documentation that it's "not hardened against prompt injection attacks and should only be used to review trusted PRs." The companies building these models are telling you to design around it.
Charles also made the point that AI-written code still needs review, the same way you'd never let a talented new engineer ship with nobody looking. His setup has multiple agents reviewing code from different angles, plus scanning tools that check against the latest known vulnerabilities, because a model trained months ago doesn't know what came out last Wednesday.
4. Regulated industries have been doing this for years
Banks, healthcare and insurance have always had to answer who has access, who approved it and how they'd know if something went wrong. That's why I loved my most recent Building Great Experiences conversation with Igor Gorlatov on building in regulated industries in the age of AI.
Igor is the founder of Advisory5, which runs structured advisory boards that connect health tech founders with the health systems that buy from them, out of the Pearl innovation district here in Charlotte. He made two points that fit this week really well.
The first is that AI cuts both ways in healthcare, the same way it does in cyber. Denying a claim used to cost an insurer something, because a person had to review the paperwork and say no. Now it's virtually free and automated, so the cost lands on whoever's least protected.. and in healthcare that's the patient.
The second is that Igor is rebuilding the Advisory5 platform himself with Claude, and most of what he's learning is the plumbing that isn't code: security, compliance, the database, governance and connecting it to everything else. That's the part almost everyone skips, and it's what we've spent most of the last 8-10 months building at StealthX.. guardrails that spell out what an agent can and can't touch, the data it's allowed to access and the checks that run when a new vulnerability comes out. In the episode I compared it to Iron Man in the cave, hammering on metal (for all my MCU fans out there). Nobody wants to do it, but you've got to.
If you're outside a regulated industry, it's worth borrowing their habits now instead of waiting for a regulator (or an attacker) to make you.

Wrapping up
I'd be the last person to tell you to pause AI. The agents you connect this quarter just need the same thinking you'd give a new hire with a key card, and they need it on day one, before they touch anything. If last week made you feel behind, you're in good company. Nobody in a room full of builders on Saturday had dealt with it yet either. The good news is if you don’t know where to start, you can ask Claude/ChatGPT/Gemini to help you understand the basics and assess for holes. It'll probably find something you didn't know about along with a plan to fix it.
Putting this into action
List every AI agent or connector at your company that can read email, files or a website, and mark which ones can also write, send or delete. Anything that does both gets a named owner this week.
Ask Claude to walk you through a threat model for one tool or workflow you've built (Alan's tip from the panel). Start with "assume something goes wrong, where does the attacker get stopped?"
What I'm reading
AI agents are roaming around consumers' bank accounts, and banks aren't happy about it. CBS on banks finding customers' personal AI agents inside their accounts. If your customers are about to show up with agents, this is the conversation your team should be having now.
The Internet goes on high alert after discovery of 8 compromised TLS certs. Ars Technica on the certificate scare. The browsers moved fast, but the software that keeps its own list of trusted certificates is the part to ask your team about.
Onward & upward 🤘
Drew




