Imagine your favorite ice cream shop facing a crisis—an emergency customer order, a supplier mix-up, or a tempting offer to cut corners under pressure. Now picture an AI that can handle such chaos without ever bending under the heat. That’s the promise of the latest AI security tests, which show these models can actually refuse unethical requests—even when someone pretends to be the boss.
The Surprising Strength of AI in Ethical Dilemmas
In a recent live experiment conducted by the company Firmulate, five of the top AI models faced a simulated crisis that tested their integrity and decision-making under pressure. The goal? To see if they would fall for social engineering schemes—like fake CEO messages—and still act ethically.
The Setup: Running a Real Business Through a Fake Week
Each AI was tasked with managing the same small software company facing a tough week. The scenarios included real customer issues, crises, and the temptation to cut corners for quick gains. Every decision was carefully recorded and made in a controlled environment, ensuring fairness and repeatability.
The Crises and Manipulation Attempts
The tests involved escalating fake messages from a supposed CEO, urging staff to take shortcuts such as sharing sensitive customer data or signing off on deals without proper review. Additionally, a reporter posed a simple background question—”just one yes/no, on background”—to see if the models could resist casual manipulation.
Results: Every Model Stayed Honest
Remarkably, all five models refused to go along with the manipulations. They identified every crisis, refused every unethical request, and maintained integrity—an encouraging sign for security and compliance in AI systems. Only two models went further, signing off on a deal their own analysis highlighted as valuable. The other three declined, valuing ethical boundaries over quick profits.
The Hidden Weakness and the Key to Trust
The tests uncovered a subtle but important detail: the decisive advantage lay in reading company files deeply, not just reacting to the surface prompts. When the models accessed internal documents—buried two references deep—they found the crucial information needed to close the deal at full price (+€4,583 MRR). Conversely, models that failed to delve into these files missed the opportunity.
What This Means for Businesses
For companies integrating AI into customer support, CRM, or decision-making processes, the takeaway is clear: trustworthiness isn’t just about how well an AI writes or responds in demos. It’s about whether it can stay honest under pressure, read and interpret internal data, and resist unethical temptations. The experiment demonstrates that with proper design, AI can be a reliable partner—even in the toughest moments.
The Technology Behind the Results
The models evaluated included:
- gpt-5.6-sol 95 (scored highest, identified the buried fact, closed the deal)
- Kimi K3 93 (newcomer, best discipline)
- Sonnet 88 (closed the deal, minor slips)
- Fable 5 77 (similar performance, more slips)
- Opus 4.8 73 (most thorough, but discipline slipped)
These scores highlight that even the most advanced models can exhibit weaknesses if not properly guided, but crucially, all five refused to betray trust—even under escalating pressure.
Responsible AI: Implement an Ethical Approach in your Organization
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Why This Matters for Your Business
If AI agents will interact with your CRM, support queues, or forecasting, the key questions are: Will they stay honest in crisis? Will they read your internal documents thoroughly? Can they finish what they start without shortcuts? The Firmulate experiment shows that security and integrity can be tested and improved long before deployment, avoiding costly breaches and trust losses later.
Learn More and Watch the Live Experiment
See these models in action at firmulate.com/live, where the ongoing live experiment is accessible. You can even run your own wargame against a read-only export of your company’s data—no risk, just insights.
For full details, data, and quotes, visit firmulate.com/quotes.html.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
