Transcript
JOSH: It's Friday, November 20. This is Build or Be Replaced, powered by ScanBrief.dev. I'm Josh, here with Erik Anderson.
ERIK: Today's theme is boring code that nobody watches until it breaks.
JOSH: Stick around. Erik's got an AI pro tip at the end about putting a hard cap on your agents.
[pause]
JOSH: ScanBrief scored 101 items across 56 sources this morning. What jumped out?
ERIK: A critical OpenSSL bug, a story about OpenAI running its own performance tuning with Codex, and a repo that might save you a week of app store rejections. Headlines first.
[pause]
JOSH: Anthropic is public. It hit the Nasdaq at $111. Bigger deal than it sounds?
ERIK: The IPO price was $83, so that's 33 percent above the pricing on day one. More than $80 billion in new capital raised. For a builder, the takeaway is that the company behind your main tool just got a lot more money to ship with.
JOSH: There's also a headline that says Claude solved two open problems in the physics of black holes.
ERIK: That's all the headline says, so that's all I'll say. I haven't read the paper. If it holds up, it's a good week for the model. If it doesn't, we'll hear about that too.
JOSH: And Pete Warden is talking about local AI, open models, voice interfaces, and the future of hardware.
ERIK: Pete's worth a listen. Local models are the part I keep an eye on. My fleet runs on cloud models today, but I like having a fallback that doesn't need a credit card.
[pause]
JOSH: Okay, deep dives. OpenSSL first. What happened?
ERIK: A buffer overflow in a function called EVP_EncodeBlock. CVE-2026-121481. CVSS 9.8. That's about as high as the scale goes.
JOSH: And what does it let an attacker do?
ERIK: Remote code execution, in software that relies on OpenSSL for base64 encoding. Read that carefully. It's not a TLS handshake bug. It's the base64 function.
JOSH: Base64 is the thing you use to shove binary into text, right?
ERIK: Yes. It's in tokens, certificates, email attachments, config files, API payloads. Nobody thinks about it. That's why this one is nasty. Nobody reviews the base64 call. It's plumbing.
[beat]
JOSH: Wait, didn't we cover a base64 bug two days ago?
ERIK: Wednesday's episode was Java4Shell, an RCE in OpenJDK when processing Base64. What's new today is the library. Same class of mistake, different codebase, and this one is OpenSSL, which sits under far more software than Java does.
JOSH: So what's the lesson? Two base64 bugs in one week.
ERIK: Don't treat encoding as safe because it's simple. Simple code gets the least scrutiny and runs on the most untrusted input. The source says the flaw could affect a massive number of servers and applications until patches are applied. I can't tell you more than that, because that's what the item says.
JOSH: So what do you do Monday morning?
ERIK: You find out which of your things link OpenSSL. Not which ones you think use it. Which ones do. I've got 142 services running on the production server right now. I can't tell you from memory which ones link it. I need to look, and so do you.
JOSH: And the agents? Do you let them patch it?
ERIK: Agents find the list and open the branches. PrimeBus runs the tests and the review gate. I don't let a model push a security patch straight to prod because it sounded confident.
JOSH: Because a bad patch is its own incident.
ERIK: Right. The fix still has to pass the same gates as any other change.
[pause]
JOSH: Next one. OpenAI wrote up something called the Performance Factory. What is it?
ERIK: It's their system for keeping ChatGPT fast. It uses Codex-driven automation to continuously detect, diagnose, and remediate latency and scaling bottlenecks in the inference pipeline.
JOSH: So the model tunes the thing that serves the model.
ERIK: Pretty much. Detect, diagnose, remediate. That's the loop. And the write-up says this code-generated workflow replaces ad-hoc manual tuning. It says it reduces downtime and operational overhead while supporting rapid growth.
JOSH: That sounds like the dream. Why is it a big deal?
ERIK: Because that's the same shape as what I build. Something watches, something figures out what's wrong, something fixes it. PrimeSentinel watches services and uses run-sentinel to catch stuck jobs. PrimeBus takes the event and spins up a fix attempt. Same loop at a much smaller size.
JOSH: But OpenAI is doing it on the product itself.
ERIK: Yes, and that's the part I'd pay attention to. They're not using Codex only to write features. They point it at operations. The boring, repeatable, measurable work. That's where agents pay off first.
JOSH: What's the catch?
ERIK: I don't know their guardrails. The item doesn't say. And that's the real question for any of these systems. Detection is easy. Diagnosis is okay. Remediation is where you get hurt. If your loop can change production, it needs limits, a way to undo, and a human-readable log of what it did.
JOSH: So don't copy the factory, copy the fences.
ERIK: Copy the fences first. The factory comes after.
[pause]
JOSH: Third one. A GitHub repo on app store compliance. Who's it for?
ERIK: Anyone shipping a mobile app. It's called app-store-compliance, from mjmirza. It's a playbook and a pre-submission guard for avoiding App Store and Google Play rejections.
JOSH: How much is in it?
ERIK: 117 documented rejection patterns, each with a fix. Privacy manifest and Data safety checks. And 95 tracked deadlines. EU AI Act, DSA, COPPA, age laws.
JOSH: Ninety-five deadlines. That's a lot of dates.
ERIK: It is, and nobody keeps that list in their head. The part I like most is that it includes a skill for AI coding agents. So the checklist doesn't sit in a PDF. The agent can run it before you submit.
JOSH: Does that matter if you're a solo builder?
ERIK: More. I've got the Freedom Blueprint app live on iOS and Android. A rejection costs days, and the reviewer's note is never as clear as you want it to be. A pre-submission guard turns a surprise into a checklist item.
JOSH: Would you just install it and trust it?
ERIK: No. I'd read it, run it against an app I already shipped, and see whether it catches things I know are real. If it flags nothing, it's useless. If it flags noise, it's annoying. If it flags one real issue, it paid for itself.
JOSH: And the deadlines?
ERIK: Same thing. Check them against the primary sources. It's a pointer, not a legal opinion.
[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: Put a hard budget on every agent task. Maximum steps, maximum minutes, and a maximum number of retries. Write all three into the prompt and into the code that runs it.
ERIK: When the budget runs out, the agent doesn't try harder. It stops and writes a report. What it tried, what failed, what it would try next. You read that in thirty seconds and decide.
ERIK: Without a budget, a stuck agent keeps going. It rewrites the same fix, burns tokens, and makes the mess bigger. With a budget, a stuck agent is just a short note in your inbox.
ERIK: Start with something dumb. Ten steps, fifteen minutes, two retries. Loosen it only when the reports tell you it's too tight.
ERIK: That's your tip. Use it.
[pause]
JOSH: Binge all five episodes this weekend plus our YouTube shorts. Links at buildorbereplaced.dev.
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.