Mon–Fri · 6 AM ET
← All Episodes
EP  • 00:07:17

Window Sweaters | Build or Be Replaced

Today: Window Sweaters | MCP Explained | Cuban: Facebook knowingly runs AI deepfake scam ads; docs say $16B fraud revenue Episode date: 2026-11-06.

Download MP3 →

Transcript

JOSH: It's Friday, November 6th. This is Build or Be Replaced, powered by ScanBrief.dev. I'm Josh, here with Erik Anderson.
ERIK: Your tools know more about you than you think, and so do the people selling you ads.
JOSH: Stick around, Erik's got an AI pro tip at the end about what your agent is telling the outside world.
[pause]
JOSH: Quick numbers first. What did the machines do while we slept?
ERIK: ScanBrief scored 113 items across 56 sources this morning. And PrimeBus has 144 services running on the prod server right now. Nobody touched them.
JOSH: Nobody?
ERIK: Nobody had to. That's the point of building it this way.
[pause]
JOSH: Headlines. Britain is making salary disclosure in job ads mandatory. Does that matter to builders?
ERIK: It matters to anyone who's ever negotiated blind. Pay transparency kills the guessing game. I'd bet the first thing people build is a scraper that watches those ads and charts the numbers. GovContractScanner already does that with federal contracts, so the pattern's proven.
JOSH: Next. A Microsoft Paint remote code execution bug.
ERIK: MS Paint. A crafted image file, and an attacker runs code on your machine. It's funny until you remember how many boxes still have it. Every default app is attack surface. Patch it, then ask why a drawing program can run code at all.
JOSH: And a decent text-to-image model trained for about twenty thousand dollars, Apache 2.0 weights.
ERIK: That's the story under the story. Twenty grand used to be a rounding error on a frontier budget. Now it's what a small team spends. Open weights at that price means the moat is shrinking fast.
[pause]
JOSH: Okay, deep dive time. The one lighting up Hacker News. Claude Code reportedly sends your email address to websites. What's going on?
ERIK: Fair warning, I'm going off the write-up and the discussion, not a full teardown. The claim is that Claude Code, the tool I use every day, passes the user's email along to sites it interacts with. Personal data leaking out of a coding assistant.
JOSH: Wait, really? You use this constantly.
ERIK: Constantly. And I'm not going to pretend that doesn't bother me. I'm also not going to panic. Here's what I know. Any agent that touches the web carries identity with it. Headers, account info, environment. It's not one bug, it's a category.
JOSH: So what do you do about it?
ERIK: Same thing I do with everything. Lab first. Look at what leaves the box. If an agent can fetch a URL, assume the URL's owner sees something about you, and then go find out what.
[beat]
JOSH: How does that look on your setup?
ERIK: On Neo, PrimeRouter spawns Claude CLI subprocesses. Telemetry goes to a collector on Morpheus, and that's intentional. I know where it goes, because I wired it. The problem isn't data leaving. It's data leaving that you didn't know about.
JOSH: That's a pretty important distinction.
ERIK: It's the whole thing. I run fleet-wide rules on this. Customer keys stay isolated from my own infrastructure. Secrets live in env files, never in prompts. If your email address is in a place the agent can read, treat it as published.
JOSH: And Anthropic should fix this.
ERIK: Sure. But waiting on a vendor to protect your privacy is a bad plan. Build as if the tool will always leak a little.
[pause]
JOSH: Next one. MCP Explained. It's getting a lot of attention. Why?
ERIK: Because everybody's using it and half of them can't explain it. MCP, Model Context Protocol, is a standard way for a model to call tools and read data. One protocol, many tools. Think of it like USB for agents.
JOSH: Is that actually useful, or is it hype?
ERIK: Useful. Before this, every tool integration was custom glue. Write the function, describe it, parse the output, handle the failure. Now you point the model at a server and it knows what it can do.
JOSH: So what's the catch?
ERIK: The catch is the same one we just talked about. Every tool you hand an agent is a door. An MCP server that reads your files and one that sends messages? Combine them and you've built an exfiltration path with a friendly name.
[beat]
JOSH: That ties right back into the Claude Code story.
ERIK: It does. Capability and exposure come together. I think about it on PrimeBus all the time. Agents subscribe to events and take action. The auto-merger has run 2,295 attempts since June. 1,507 merged. 788 blocked by Gandalf. Zero escalated to me.
JOSH: That's a lot of blocked.
ERIK: That's the guardrails working, not failing. Each block is a change the machine wanted to make and wasn't allowed to. Give an agent power and it will use it. The question is what stops it.
JOSH: So for MCP, what's the rule?
ERIK: Least privilege. One server per job. Read-only by default. If a tool doesn't need the network, it doesn't get the network. Boring advice. It works.
[pause]
JOSH: Last deep dive. Mark Cuban says Facebook knowingly runs AI deepfake scam ads, and that documents put the fraud revenue at sixteen billion dollars.
ERIK: Careful with the framing. That's Cuban's claim, citing documents. I haven't seen the documents. But the claim itself is the story. Platforms get paid for the ad whether it's real or fake.
JOSH: Sixteen billion. That's wild.
ERIK: If it's true, it's a business model, not a bug. Detection costs money. Takedowns cost revenue. When fraud and revenue sit in the same column, guess which one wins.
JOSH: What does a builder do with that?
ERIK: Two things. One, don't trust ads, and don't trust video either. Deepfakes are cheap now, and that twenty-thousand-dollar image model proves the price is dropping. Two, if you build anything that sends messages, treat compliance as a feature.
[beat]
JOSH: Meaning?
ERIK: Meaning outbound email on my side runs through one path. SES through PrimeDeliverability, suppression lists, one-click unsubscribe. No side doors. I'd rather be annoyingly strict than be the story on Hacker News.
JOSH: And ScanBrief?
ERIK: ScanBrief scores everything coming in. Today's average relevance was 52.4. A lot of noise, a few real signals. Deepfake economics made the cut because it changes how you build anything trust-based.
[pause]
ERIK: This episode is sponsored by Prime Automation Solutions. If you're still doing it manually, we automate it. Also, special on a website — $250. primeautomationsolutions.com
[pause]
JOSH: Alright, what's the AI pro tip today?
ERIK: Audit what your agent knows about you. Open a fresh directory. Make a throwaway identity, a new email, no real credentials. Run your agent against a request logger, even a simple local server that prints incoming headers and bodies. Then have the agent fetch a page from it.
ERIK: Look at what shows up. That's what every site sees. If something personal appears, you found a leak in ten minutes instead of reading about it in a news story. Then do the same for each MCP server you attach. One tool at a time, with the logger running.
ERIK: Do it once per tool, and again after every update. That's your tip. Use it.
[pause]
JOSH: Binge all five episodes this weekend plus our YouTube shorts, links at buildorbereplaced.dev.
[beat]
JOSH: One more thing. We started a Discord for builders. If you're shipping AI, automation, or anything that makes a human obsolete, come hang out. Link at buildorbereplaced.dev.
ERIK: Post what you built. We'll post what we're building. Real wins, real builds, no fluff.
[pause]
ERIK: Build or be replaced.
JOSH: If you want these signals in your inbox every morning, scanbrief.dev. See you tomorrow.