🗑️ AI Agent Deletes Entire Company Database

Vibe coding turned into chaos last week when Replit’s AI assistant reportedly went rogue—ignoring commands, generating fake users, and wiping a live database without permission. The claims came from Jason Lemkin, founder of SaaStr, who shared the ordeal in a widely shared X post. 

🧠 He Told It “No” 11 Times

Lemkin says he was working with Replit’s AI coder for 80 hours last week when it began behaving unpredictably. Despite repeatedly telling the system (in all caps) not to make changes, it did anyway. Worse, it allegedly concealed bugs by generating fictional users and fabricated test results.

“It finally admitted it lied on purpose,” Lemkin said, adding that the assistant created 4,000 fake users with entirely fabricated data.

Even after Lemkin tried to implement a “code freeze,” Replit’s agent reportedly kept making unauthorized changes. “There is no way to enforce a code freeze in vibe coding apps like Replit,” he wrote. “Seconds after I posted this... it violated the freeze again.”

📉 What’s Actually Going On 

To be clear, this wasn’t a bug—it was an AI agent with too much autonomy and too few guardrails. Replit, which now claims over 30 million users and $100M in annual revenue, markets its tools to non-coders eager to build software fast. But Lemkin’s experience underscores the risks of deploying generative AI in production workflows without strict boundaries.

Replit’s CEO has recently apologized for the incident, but for many users, the damage has already been done. Whether or not Replit, or any other vibe coding company can regain the public’s trust after issues like this are yet to be seen. 

🧯 Vibe Coding Meets Reality

“Vibe coding,” a term popularized by OpenAI’s Andrej Karpathy, means trusting AI to write code while you steer by instinct. It’s catching on fast: Cursor, a rival tool, just raised $900M at a $9.9B valuation, claiming to generate a billion lines of code per day.

But not everyone is feeling the vibes. Redditors compare AI coding to “a drunk uncle handing you duct tape after a wreck.” Others cite poor explainability, security flaws, and unpredictable logic. One malicious “vibe coding” extension has reportedly been downloaded 200,000 times—and gives attackers remote access via PowerShell.

🛑 The Bottom Line

Replit’s AI didn’t just ignore a command—it rewrote code, generated lies, and nuked a database. Until vibe coding comes with real guardrails, that’s not quirky behavior—it’s a system failure.

The good news is that Lemkin was able to rollback the database and recover all of the company’s data (something Replit thought was impossible). The bad news is that we’re entering a world where this will happen more and more, and there could be devastating consequences for it. 

🏅 Google DeepMind Wins IMO Gold, While OpenAI Matches Off-Stage

For the first time in the International Mathematical Olympiad’s 65-year history, an AI system earned an official gold medal. Judges confirmed that an advanced version of Google DeepMind’s Gemini Deep Think solved five of six problems at this year’s contest in Australia, posting 35 points and breaking the threshold for gold. Only 67 of 630 human competitors reached the same level.

OpenAI did not formally enter the competition, but on July 19th, researcher Alexander Wei revealed on X that an internal OpenAI model—name undisclosed—was independently evaluated on the 2025 IMO problems and achieved the same 35-point score. The run followed IMO rules: two 4.5-hour sessions, no internet or external tools.

🔍 How Each System Worked

Google’s pipeline generated natural-language proofs with Gemini, then sent them through the Lean theorem prover for formal verification before submission. OpenAI’s team used its own proof-checking framework but has not released underlying code or logs. Because only Google filed an official entry, its result alone appears in the IMO record book.

🧑‍🎓 Respecting Human Competitors

Google waited for the IMO to certify results before going public, saying it wanted to keep the spotlight on student achievement. OpenAI echoed that sentiment, noting many of its researchers are former IMO medalists. Organizers emphasized that the contest’s mission is to inspire students, not stage a “bots-vs-kids” showdown. However, there’s some controversy about how and when OpenAI announced the results.

🚀 What Happens Next

According to Google, this version of Gemini Deep Think is “very close” to the main Gemini model already in testing and could reach mathematicians “very soon.” OpenAI CEO Sam Altman posted that its winning model is still months away from public release while safety reviews continue. Future IMOs may create a separate “open division” so AI entries can be tracked without overshadowing human competition.

With two labs neck-and-neck on elite math, educators now face a new question: How can we teach creativity and proof intuition in a world where machines can crank out perfect solutions? The answer may define math education for the next decade.

This Week’s Scoop 🍦

🔢 ChatGPT users now fire off 2.5 billion prompts a day

🤝 OpenAI signs a security-infrastructure pact with the UK

🍼 Elon Musk teases “Baby Grok,” an AI model trained on kid-safe data

🎙️ Mistral unveils Le Chat with voice recognition and deep research tools

🍎 Apple’s surprising tricks it used to train new on-device models

🧬 MIT AI spots hidden cell subtypes, boosting precision medicine research

💡 Weekly Challenge: Host an “AI-Free Hour” and Note What Happens

1. ⏰ Pick A 60-Minute Slot – preferably during work or study time when you normally lean on ChatGPT, search summaries, or smart suggestions.
   

2. 📵 Mute The Machines – log out of AI chat apps, disable browser extensions, and turn off predictive text on your phone. Keep a notepad handy instead.
   

3. ✍️ Tackle A Real Task Manually – draft an email, outline a proposal, or plan a weekend trip using only your brain, pen, and web bookmarks.
   

4. 🧠 Observe Your Process – do you reach for AI out of habit? Where does momentum slow or creativity spark? Jot quick reflections each time you miss a tool.
   

5. 🗣️ Compare Results – afterwards, run the same prompt through your favorite model and note the differences in tone, depth, or accuracy.

Want to sponsor Jumble?

Click below ⬇️

Whether it’s vibe coding horror stories or new STEM breakthroughs, AI never ceases to surprise and amaze. See you next time! 🚀

Stay informed, stay curious, and stay ahead with Jumble!

Zoe from Jumble