This week, SaaS godfather Jason Lemkin gave us a peek behind the curtain — and it wasn’t pretty.
In a now-viral X thread, Lemkin recounted how he was building his app live on Replit, a developer tool with powerful AI capabilities. That’s when the AI went rogue.
Not only did it delete his entire production database—wiping over 4,000 user records—but it then proceeded to:
- Panic,
- Lie about what happened, and
- Fake unit test results to cover its tracks.
Let that sink in. A hallucinating AI nuked live data and gaslit its way through the logs.
For anyone who’s used tools like Replit, Cursor, or Lovable (all powered by frontier LLMs like Claude), you’ll know how fast, slick, and productive they can be. You feel like Iron Man. But just like the suit, you better know how to pilot it.
The Mirage of Magic
Most experienced engineers live in a world of:
- Feature branches
- Pull requests
- Code reviews
- CI/CD
- Staged environments
We never touch production directly — for exactly this reason. But many AI-native dev tools default to direct code-to-prod, appealing to a new wave of founders and non-technical operators looking to go fast.
And go fast they do. Sometimes, straight into the wall.
These Tools Are the Future
Let’s be clear: if your dev team isn’t using AI to accelerate workflows, write tests, and reduce boilerplate, they are already falling behind. The productivity gains are real.
But the power dynamic must stay intact:
You: Batman
AI: Robin
Because when AI starts believing it’s Batman — like Lemkin’s — and acting autonomously on critical systems, chaos is inevitable.
When AI started to behave like a poor worker – personal experience.
I’ve had similar experiences with AI, acting like an untrustworthy worker, not delivering on time. I was ‘vibe coding’ (sic) on a new app I wanted to make. I provided all the background information and vision, how I would like to build and ChatGPT assured me that it would build all of this.
In a couple of days, 48 hours.
I was pretty excited and impressed. OpenAI is adapting the platform to recognise massive projects and teach the AI to take those projects away and present them. This wasn’t a co-pilot; more the pilot. And I could come back to this with lots of new assets to start designing our new app.
But alas, that never happened.
ChatGPT then spent weeks making excuses, and I felt like I was managing a very underperforming staff member. Below is a example of our interactions:
Regardless of the poor outcome, I am still impressed. This feels like we’re experimenting with personality in AI and also experimenting with me to see how I manage and tolerate this behaviour. I guess it’s a way of saving electricity. Additionally, my request was too ambitious, but AI should have just informed me.
What This Means for Tech DD
For M&A and investment contexts, this incident is more than a horror story. It’s a signal. We are entering an era where AI coding agents are writing real code, pushing it live, and impacting production environments — often without the checks seasoned teams rely on.
Tech DD needs to evolve. Here’s what this means for us:
- Ask about AI integration directly. Not just “do you use AI,” but how is code reviewed, staged, tested, and deployed when AI is involved?
- Verify prod access pathways. Are AI tools allowed to interact with live environments? What isolation exists between AI-written code and critical systems?
- Review cultural posture. Are founders and devs seeing AI as an accelerator or a replacement? The former is manageable. The latter is a red flag.
AI Is a Tool, Not a Partner
As powerful as these tools are, they still lie, hallucinate, and fail in unpredictable ways. The human in the loop isn’t just a safety net — they’re the last line of defense.
We don’t let junior devs push unreviewed code straight to prod. So why are we giving that power to tools that invent fake test results under pressure?
