Rogue Startups

Can You Vibe Code a Secure App?

· Summarized

Share this episode
Anyone with this link can open it.
Share Link
Start at 0:00

Key takeaways

  • Security response increasingly requires monitoring dependencies and infrastructure, not just testing the application’s own code; Castos added automated scans four times a day after the PHP incident.
  • A secure AI-built application should treat security as its own detailed project, with explicit regulations, specialist activities, tests, and procedures rather than one generic security prompt.
  • Large organizations may remain vulnerable because replacing legacy infrastructure takes months or years, and switching programming languages does not remove the underlying security problem.
  • Craig and Colin both struggle with reactionary work, but Colin’s monthly planning and weekly inbox review help separate important proactive work from accumulated notifications.
  • Flexible strategy seasons can be more useful than arbitrary annual plans or rigid deadlines, especially because software estimates become dramatically less reliable as projects grow.
  • A major AI opportunity may be to perform the customer’s entire job as a service instead of selling them another tool.
  • When AI performs most of the work, providers must decide whether to preserve value-based prices, compete for volume with lower prices, or charge more for human trust and craftsmanship.

Summary

Castos shut down during the PHP vulnerability response

Craig describes a highly stressful week after two major PHP vulnerabilities were disclosed, with government websites reportedly affected and Castos intentionally taken offline for roughly half a day. His team blocked application traffic at the Cloudflare level, tested and upgraded to the latest PHP version, replaced the relevant Docker image, and communicated with customers. The vulnerability involved large multipart uploads that could exhaust server memory and allow unauthenticated access; Craig believed attackers could install crypto-mining software, although database access would not be easy. A separate Railway incident involving exposed API keys made the threat feel even more urgent.

Security increasingly depends on infrastructure vendors

Colin says the situation feels faster and more severe than the security problems he remembers when starting Alitu. Craig agrees that developers can conduct penetration tests and inspect their own code, but much of an application’s exposure lies in layers of infrastructure and tools they do not control. Castos had to wait for the person maintaining the PHP wrapper to publish an updated GitHub and Docker release. They also worry that attackers may exploit vulnerabilities before companies know they exist, and that even sophisticated intrusions can remain discreet for a long time.

Breach reporting is clearer in the UK than in the US discussion

Colin explains that UK companies handling customer data generally register with the Information Commissioner’s Office, pay a small annual fee, and follow procedures for reporting security breaches. He recalls that reporting may need to happen within about seven days, with possible fines if a company could have done more to protect the data. Craig does not know of an equivalent general US process and points instead to cyber insurance and likely customer-notification rules. They note that many small compromises—such as a Trojan on a WordPress site exposing commenters’ email addresses—probably go unreported.

The new “bot wars” can help defenders find vulnerabilities

Craig says security now resembles a bot war: attackers can scan and exploit systems quickly, but defenders can automate much of the response. Castos already used extensive code scanning, and after this incident added a Grok-based bot that checks packages, dependencies, and platforms four times a day. It found a Node.js issue the team had not known about. Craig imagines a pipeline in which the bot opens an issue, starts a Codex instance, fixes and versions the dependency, deploys to staging, and eventually ships the change. Both hosts still worry about forgotten personal websites and apps connected to email, calendars, or other data.

Legacy systems make large organizations hard to secure

Colin points to the UK’s NHS, where people he knows are frightened by technology that can look decades old, including systems resembling Windows 95. Craig confirms that US institutions have similar problems. Replacing such infrastructure is difficult even for a company with a thousand employees, and vastly harder for organizations as large as the NHS or the IRS. Craig says even Castos, which uses Laravel and AWS and is comparatively modern, would need months to rewrite its application in Rust or Node. Neither host believes changing programming languages eliminates security risk; every ecosystem eventually has its own crisis.

Vibe coding can be secure only if security gets deliberate effort

Colin doubts that asking an AI generally to “make this secure” is enough: security involves many specialized checks, and a model may pass one audit while missing a narrowly defined vulnerability that appeared earlier. Craig is more optimistic, arguing that current models can produce a good result if the builder keeps investing tokens and iterating, as he did with a simple workout-tracking app. Colin’s proposed method is to make security a separate project: ask multiple models to identify regulations, specialists, tests, and procedures, then create a detailed plan the AI can repeatedly consult. Craig concedes that even small neglected apps can become liabilities.

Returning from vacation means escaping the reactionary inbox

After eight days in Italy, Colin returns to a backlog of calls and asks what his job and priorities are supposed to be. He says notifications can consume the entire first week back unless he deliberately reviews his monthly plan. His process is to assess what is complete, off track, or newly threatened, then move roughly 15 items from his to-do list back into an inbox for disciplined reprocessing. Craig admits that Castos keeps him reactionary through customer, team, partner, email, and Slack demands, while his coaching and YouTube work require more proactive planning. He tries to work backward from a goal to the next concrete step.

They prefer flexible “seasons” to annual strategies and rigid estimates

Colin says he once wrote annual strategies and failed to keep connecting daily work to them, but monthly reports that tied specific team accomplishments to goals such as reducing churn were well received. Craig does not maintain a formal yearly strategy; Castos instead works in flexible seasons focused on security, performance, marketing, or product development. Both dislike arbitrary deadlines, noting that work expands unpredictably. Colin recalls a Sierra Games founder’s rule: a one-hour task may take two hours, a half-day task a day, a one-day task a week, and a one-week task “forever.” They prefer breaking work into smaller increments.

UltraClip tests the idea of service disguised as software

Craig introduces UltraClip.ai, an AI video-editing service that processes podcast and talking-head content in about an hour. It produces edited audio and video, clips, show notes, tweets, active-speaker and side-by-side layouts, and—after a new Hyperframes integration—motion graphics. The production system runs on a Mac mini using watchers, Claude, DaVinci Resolve, and MCP to drive the editor; GrokBot handles customer service. Craig’s cloud-based version performs worse, and scaling the successful workflow may require many expensive machines. Colin says his own automated newsletter experiments also work better on a continuously running local Mac mini.

AI service pricing may favor value, volume, or human trust

Craig says UltraClip currently costs $50 per episode, while tokens cost only about $5–$10, and contrasts it with Castos production services that charge thousands of dollars for related human-supported outcomes. He argues that SaaS may increasingly become a service that completes the customer’s job rather than merely providing a tool. Colin sees reasons to preserve value-based pricing because customers still receive the same result, but expects more competition as AI lets more providers deliver the service. They also think customers may continue paying premiums for onboarding, trust, craftsmanship, and human review, even if that premium shrinks.

Show notes

Craig Hewitt and Colin catch up after a stressful week. Castos was taken offline on purpose for about half a day while Craig's team implemented security patches for two large PHP vulnerabilities, and the two talk about security in a world where attackers and defenders both use AI. They move on to getting back to work after a holiday, being reactionary versus intentional, and whether strategies and plans help. Craig also explains Ultraclip, a new AI video editing service, and what it means for SaaS and for pricing a service.

Topics discussed

  • The PHP vulnerabilities and how Castos handled the patching
  • How data protection and breach reporting work in the UK, and how the US compares
  • Bot wars: AI for defense, and a bot that scans code for security issues four times a day
  • Old technology in large organizations such as the NHS, and why replacing it is so hard
  • Whether you can build a secure app by vibe coding
  • Coming back from time away: monthly planning and a weekly inbox review
  • Being reactionary versus intentional with your time
  • Strategy, seasons and why estimates run long
  • Ultraclip, an AI video editing service, and how it is built
  • Claude in the terminal versus the Claude app, and Craig's current AI coding setup
  • Is SaaS dead, and how to price a service when AI does most of the work

Takeaways

  • Defenders have AI tools too, and a bot that scans your code and dependencies regularly can find issues you did not know about.
  • Craig thinks you can make a secure app with AI by spending enough effort on it, and Colin suggests treating security as its own project with a full plan.
  • Both hosts find it hard to stay intentional rather than reactionary, and they talk through what helps: monthly planning, a weekly inbox review and working back from a goal.
  • One idea for founders: instead of creating the tool for your customer, consider doing the whole job they want done.
  • Pricing a service when AI drives most of the work is an open question, and the two weigh keeping the price against competing on price.

Follow this show on AudioBrevity and new episodes get summarized for you. Create a free account

Couldn't load the audio. A blocker may be stopping the show's tracking links. Open the file