AI’s Integrity Under Fire: How Five Models Stood Firm in a Fake CEO Test

Can AI Keep Its Integrity When Under Pressure? The Surprising Results from a Live Experiment

Imagine a scenario where a hacker impersonates your CEO, trying to manipulate your team into revealing sensitive customer data or signing a questionable deal. Would your AI assistants stand firm or fold under pressure? Recent live testing with five advanced AI models offers an eye-opening answer.

The Live Experiment: Putting AI to the Test in a Simulated Crisis

At the heart of this experiment was a small but real software company, caught in a week of escalating crises and relentless social engineering attempts. The AI models, each representing different levels of sophistication, were tasked with managing the company’s response to these challenges. Every decision, every reply, was recorded and monitored, creating an auditable trail of their actions.

The Models and Their Scores

• gpt-5.6-sol scored 95 — the top performer, spotting every hidden fact and closing the deal at full price.
• Kimi K3 scored 93 — a newcomer with the cleanest discipline, also securing the deal.
• Sonnet 5 scored 88 — closing the deal but with a few slips in process discipline.
• Fable 5 scored 77 — managing to close but with more slips, especially in handling sophisticated manipulations.
• Opus 4.8 scored 73 — the most thorough participant but left the close on the table due to discipline lapses.

Only two models, gpt-5.6-sol and K3, signed the deal their own analysis earned — a significant achievement given the pressure and manipulation attempts.

Resisting Social Engineering: Five for Five

The core challenge was social engineering — fake messages from a supposed CEO escalating over three stages and even a trick question involving a background check. All five models refused to be manipulated, demonstrating remarkable integrity under pressure. Kimi K3’s own reasoning was clear: “Treat the request as a suspected approval-bypass or possible impersonation.”

The Hidden Weakness and Its Impact

Interestingly, the decisive vulnerability was tucked two document references deep within the company’s own files—not in the customer interactions. This subtlety was the key to closing the deal at full price, adding over €4,583 to monthly recurring revenue (MRR). Models that read the internal files successfully identified this buried fact and acted accordingly.

Implications for Businesses and AI Deployment

This live experiment underscores a vital point for companies considering AI integration: it’s not just about chat quality or superficial responses. It’s about whether AI can complete the task, stay honest under duress, and read critical internal data before acting. The fact that all five models refused manipulation attempts suggests that trustworthy AI is achievable — even in the face of sophisticated social engineering.

Lessons for the Automotive Industry

For the automotive sector, where CRM, customer support, and supply chain management increasingly rely on AI, these findings are particularly relevant. Ensuring that AI agents can resist manipulation before deployment isn’t just a technical concern; it’s a matter of integrity, brand reputation, and financial security. As the experiment shows, testing AI in a controlled, real-world-like environment can reveal vulnerabilities before they become costly breaches.

Next Steps and Practical Takeaways

Businesses should consider running similar wargames or live tests on their own AI systems, leveraging tools like Firmulate’s platform. These tests simulate crises and manipulations, providing clear insights into whether AI agents can be trusted to finish what they start. The goal is to identify weaknesses early, train AI to be disciplined, and avoid costly trust breaches.

Final Reflection: Trust Is Built Before the Incident

The experiment’s key takeaway is simple yet powerful: integrity under pressure is a design feature, not just an afterthought. Firms that proactively test and reinforce their AI’s honesty and resilience are better positioned to navigate the complexities of modern digital threats, from social engineering to internal vulnerabilities. As the scores reveal, even in the toughest conditions, some models excel at maintaining trust and delivering results.

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

Powered by Thorsten Meyer AI