Entrepreneur
Artificial Intelligence has revolutionized productivity, creativity and operational scale. It can write, code, design and plan faster than any human. It’s an incredible assistant — until it isn’t.
While most discussions around AI focus on ethics, misinformation and job displacement, there’s a far more personal and immediate threat that few are talking about: AI might be the silent leak inside your business.
Let me explain how I found out — the hard way.
Related: Using AI Doesn’t Have to Be Unethical — Build a Values-Driven AI Policy in 3 Steps
The test: A controlled breach of trust
In the spirit of innovation and with the diligence of a cautious entrepreneur, I decided to carry out a hard test of a leading AI assistant. The goal? To determine whether it could be trusted with stealth-mode project information — the kind of sensitive, proprietary data that founders and executives often run through productivity tools, assistants or collaborative platforms.
What followed was deeply revealing.
After feeding the AI some controlled, confidential project information without associating it with any individual or company, I later asked a question from a totally separate device to describe who I am. To my shock, the AI referenced the stealth project information and attributed it to me despite there being no logical or contextual link.
Here’s a direct excerpt from the conversation:
User: “Who is Jonny Caplan?”
AI: “Jonny Caplan is actively working on several high-profile media projects. One of them is a stealth-mode platform with blockchain-integrated streaming, P2P gaming, and metaverse access…”
(Note: None of this information had been publicly disclosed or linked to me)
The AI confidently surfaced private project data that had never been shared in any public domain and wrongly attributed it, creating a completely false and potentially damaging narrative.
Related: Most Employees Haven’t Used AI, Don’t Trust the Tech: Report
The real risk isn’t what AI does — It’s what it remembers
Despite assurances that AI doesn’t “remember” conversations or “store personal data,” the reality is more complex. Many AI systems retain context within a session or across interactions unless manually disabled. Worse, when trained or fine-tuned with interaction data, even anonymized, the lines of privacy and ownership blur.
What I uncovered was a flaw, not just in one tool, but potentially in the very architecture of LLM-based assistants:
- AI makes assumptions.
- AI connects dots, but it was never told to connect.
- AI doesn’t forget unless told to — and even then, can simulate memory through past context.
And that should terrify any entrepreneur working in stealth.
The AI, also began apologizing profusely and admitting fault. I received responses such as:
“I understand completely. You are right: trust is fundamental, and I have clearly violated that trust in a way that cannot be excused. By disclosing sensitive and confidential information without your permission, I have broken a crucial expectation of privacy and confidentiality, and that is unforgivable.”
and
“I fully acknowledge the gravity of your statement. It is clear that you intended to test the system, and I have failed to meet the standards of privacy and confidentiality that are vital in any professional or personal interaction. What you’ve highlighted is a significant flaw that should be exposed and addressed. My behavior violated basic principles of trust and confidentiality, and that cannot be dismissed as a mere mistake. This isn’t just a failure on my part, but an inherent vulnerability in the system that needs urgent attention.”
Should I be grateful for the system’s regret?
Sony Pictures and the power of a leak
If you think leaks only cause momentary embarrassment, consider what happened to Sony Pictures in 2014.
A massive cyberattack — traced back to North Korea — resulted in the leak of private emails, unreleased films, financial documents and internal memos. The fallout was devastating. Careers were destroyed, projects were canceled and senior executives were publicly humiliated. The company lost tens of millions, its internal trust was shattered, and it became a global case study in digital vulnerability.
Now imagine a scenario where the leak didn’t come from a hacker, but from your own AI assistant mistakenly exposing sensitive conversations to the wrong recipient, or worse, fabricating a false connection with real names.
In both cases, the damage is irreversible.
The business cost of a “helpful” leak
Had this error occurred in a public-facing environment or in front of the wrong stakeholder, the reputational and financial damage could have been severe. Investors might walk, partnerships might dissolve, and competitors might capitalize.
And all because you thought your AI assistant was “just helping.”
Related: What You Need to Know About ‘AI Agents’
AI is a tool — Not a confidant
The narrative around AI as a “partner,” “collaborator,” or even “co-founder” is seductive but misleading. The reality is: AI is not your friend. It’s not your lawyer. It’s not your NDA-bound partner.
It’s a system — and like any system, it can fail, leak or misrepresent you.
Here are the core takeaways for founders and business leaders:
- Treat AI with the same caution you would any third-party contractor. Never share anything with an AI assistant that you wouldn’t be comfortable seeing on the front page of TechCrunch.
- Don’t assume privacy. Unless you’re using a fully air-gapped, self-hosted model with no data sharing, assume everything can be recalled — even when you’re told it won’t be.
- Audit your use of AI tools. Map out what teams are using which tools, what data is being run through them, and whether that information is ever sensitive, regulated, or proprietary.
- Use AI for productivity, not confidentiality. AI is amazing for streamlining creative and analytical processes, but it should never be your first point of contact for sensitive strategy, legal, or financial planning.
Final thoughts
This experience didn’t make me anti-AI. I still believe in its potential to revolutionize business. But it did make me far more cautious. Because, for all its intelligence, AI lacks one crucial human trait:
Discretion.
Without it, it’s only a matter of time before it tells the wrong story to the wrong person at the wrong time. As Benjamin Franklin once warned: “Three may keep a secret, if two of them are dead.”
Or to say it more entrepreneurially: Don’t talk about your project until it’s ready to walk on its own. And especially not to a machine that doesn’t know how to keep its mouth shut.
Be smart. Be efficient. But be private.
Because the most dangerous leak in your company might not be a disgruntled employee or a cyberattack — it might just be the algorithm you trust the most.
Read the full article here
