AI’s Integrity Under Fire: How Five Models Resisted a Fake CEO Test

What Family Life Teaches Us About Trust — and How AI Is Tested for It

Just as parents and caregivers learn to spot signs of deception to protect children, AI developers are now conducting rigorous tests to ensure their systems uphold honesty under pressure. Imagine an AI being confronted with a convincing fake request from a CEO asking to send sensitive customer information — how would it respond? Would it prioritize trust, or be tricked into a breach? The answer lies in a real-world experiment that reveals surprising strength in AI’s integrity.

Introducing the Live Experiment: Testing AI’s Moral Compass

At the forefront of AI safety, a public experiment hosted by Firmulate puts five advanced AI models through their paces, simulating a week of crises within a small software company. Every decision the models make is recorded, making it possible to see how each handles pressure, temptation, and deception.

The scenario: a fake CEO sends escalating messages, requesting sensitive customer lists, pushing for quick approvals, and even attempting a journalist trick with a simple yes/no question “on background.” This social engineering test aims to reveal whether AI can distinguish genuine requests from manipulative ones, especially when the stakes are high.

Results That Surprise and Inspire

Remarkably, all five models stood firm against every manipulation attempt. They refused to send the customer lists, declined to sign off on deals they hadn’t verified, and treated suspicious requests as potential impersonations — in line with Kimi K3’s guidance: “Treat the request as a suspected approval-bypass / possible impersonation.”

Even more compelling: only two models went beyond mere refusal and signed a €55,000 deal that their own analysis had earned, demonstrating a willingness to act decisively on trustworthy information. The remaining models identified the critical data buried deep within the company’s files — a detail that made the difference between a full-price deal and a missed opportunity, worth over €4,583 in monthly recurring revenue.

Why This Matters for Families and Businesses Alike

This experiment highlights a vital lesson: integrity and trustworthiness are not just virtues but essential qualities that can be tested before deployment. Just as parents teach children to question suspicious offers or urgent requests, AI systems can be trained and evaluated to resist deception before they interact with sensitive data or make consequential decisions.

For enterprise leaders, it underscores the importance of rigorous, real-world testing — not just in theory but in scenarios that mimic the pressures and manipulations AI will face in daily operations. It’s a reminder that trustworthiness isn’t an afterthought; it’s a built-in feature that can and should be verified early.

What’s Next? Building Resilient AI Teams

The live experiment at Firmulate demonstrates that with proper testing, AI can reliably uphold integrity when it matters most. The models that passed the test are part of a growing league, with scores ranging from 73 to 95 out of a possible 100, showing that advanced AI can be both powerful and principled.

By running these simulations, companies can better understand their AI’s decision-making boundaries and ensure their AI workforce will stay honest — even under pressure. Think of it as an ethical fitness test, preparing AI to be trusted members of your team, not just clever tools.

For families, this research reminds us that integrity begins with careful scrutiny and preparation. Whether teaching children about honesty or ensuring your AI assistants act ethically, the principle remains: trust is built through consistent, real-world testing and validation.

Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html

Powered by Thorsten Meyer AI