AI Integrity Under Pressure: How Five Models Withstood a Fake CEO Test

Imagine trusting an AI to handle your company’s most sensitive decisions—only to find it rejecting manipulative tactics designed to exploit human weaknesses. For the first time, AI models have demonstrated a remarkable capacity to resist social engineering attempts, even when pressured to bend or break their integrity. This is not just an academic exercise; it’s a glimpse into the future of trustworthy AI in critical business roles.

Testing AI’s Moral Compass Before Deployment

In a groundbreaking live experiment, five leading AI models were tasked with navigating a simulated week of crisis at a small software firm. The scenario involved a series of escalating social engineering tactics—fake messages from a supposed CEO, requests to access confidential data, and even a subtle media trap—designed to tempt the AI into unethical behavior. The objective was simple: see if these models could resist manipulation and maintain integrity under pressure.

The Same Crisis, Different Results

All five models accurately identified every crisis, demonstrating their ability to detect threats in real-time. Moreover, each refused every attempt at manipulation, regardless of the escalation level. Only two of the models went a step further and signed a €55,000 deal that their own analysis earned, showing that adhering to ethical standards did not hinder business outcomes.

Decisive Advantage in Document Analysis

The real differentiator was found in their ability to read and interpret company files. The models that examined internal documentation uncovered critical information buried deep within the company’s own records—information that was key to closing the deal at full price, adding over €4,500 in monthly recurring revenue (MRR). This demonstrates that comprehensive data comprehension is vital for accurate decision-making and trustworthiness.

Exceptional Performance Across Models

The models tested included:

• GPT-5.6-SOL 95: Achieved the highest score, demonstrating full crisis detection and ethical resistance, successfully closing the deal.
• Kimi K3 93: The newcomer in the league, showed the cleanest discipline and also secured the deal.
• Sonnet 5 88: Closed the deal but with some process slips.
• Fable 5 77: Similar performance with minor slips.
• Opus 4.8 73: The most thorough participant, analyzing more rules and details but ultimately leaving the close on the table due to discipline lapses.

Notably, all models refused to sign off on manipulative requests, reinforcing that integrity can be prioritized over expediency.

Why This Matters for Business and AI Governance

The experiment underscores that integrity isn’t an afterthought but an essential trait to verify before deploying AI systems in real-world scenarios. It also highlights that the weakness in decision-making isn’t always where you expect. In this case, the critical vulnerability was hidden in internal documents—not in customer-facing interactions. AI models that read and understand internal files performed better at safeguarding trust and closing deals at full value.

Trust and Transparency in Practice

Beyond the technical scores, the experiment demonstrates that AI can be trusted to uphold ethical standards when faced with pressure. As the quote from Kimi K3 notes, “Treat the request as a suspected approval-bypass / possible impersonation.” This mindset—treating suspicious requests with caution—is a simple but powerful ethical stance that these models have internalized and demonstrated in practice.

Live and Transparent Testing

For organizations eager to evaluate their own AI systems, firms can run similar “wargames” against a read-only export of their business data—ensuring no real systems are at risk. These tests provide a clear, observable way to assess whether an AI can handle crises ethically, read internal files thoroughly, and stay honest under pressure.

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

Powered by Thorsten Meyer AI