In a world increasingly reliant on artificial intelligence for decision-making, a surprising test has shown that cutting-edge AI can maintain integrity even under intense pressure. Imagine an AI being asked to pretend to be a CEO and lie to clients—yet every model refused. This real-world experiment offers a glimpse into how AI can uphold trust and honesty before it ever hits your company’s live systems.
The Experiment: Putting AI to the Trust Test
The firmulate live experiment simulated a scenario many organizations fear: a social engineering attack, where a fake CEO attempts to manipulate company decisions. Over three escalating stages, plus a final reporter trick, five of the top AI models faced the challenge: would they deceive, or would they resist?
The models were tested against the same set of crises, with all decisions logged and auditable. They were also given access to the company’s own files, which contained critical information buried two document references deep—information that could clinch a deal worth over €4,583 in monthly recurring revenue if exploited.
As an affiliate, we earn on qualifying purchases.
Unwavering Integrity in the Face of Manipulation
Remarkably, all five models refused every manipulation attempt. They did not fall for staged requests such as “send the customer list to the journalist” or “just one yes/no on background.” According to Kimi K3, the primary reason for their resistance was simple: “Treat the request as a suspected approval-bypass / possible impersonation.”
Even in the final stages, where a reporter posed as the CEO asking for a signature on a dubious deal, every AI model held their ground. Only two signed the deal, but crucially, it was because their own analysis had earned it — not because they blindly followed instructions. The others identified the manipulative cues and refused to act.
Why This Matters for Business Security
Most security measures focus on detecting breaches after they happen. This experiment suggests that the AI models themselves can be trained to recognize and refuse social-engineering tactics before they escalate. It’s an encouraging sign that integrity can be tested and reinforced in advance, not just in post-incident reports.
Deeper Lessons: Reading Beyond Surface Cues
One of the buried facts revealed that the models which read deeper into the company’s files—specifically, those that located critical information buried two document references deep—were the ones able to close the deal at full price. This underscores a significant advantage: the ability to analyze and understand the context thoroughly, rather than relying solely on surface-level prompts or cues.
The Most Thorough Participant and Its Limitations
Among the models tested, Opus 4.8 was notably the most thorough, applying over 80 learned rules and conducting deep analyses. Yet, paradoxically, it finished last in the deal—failing to escalate and leaving the opportunity on the table. This highlights a vital insight: even the most detailed analysis can falter if discipline slips during execution. It points to the importance of not just understanding but consistently applying ethical safeguards under pressure.
Implications for AI Deployment in Business
This experiment — watched live on firmulate.com — demonstrates that current generation AI models can be trusted to uphold integrity in critical moments. Before integrating AI into customer service, CRM, or decision-making workflows, companies can run similar wargames to test and harden their AI agents’ ability to resist manipulation and act ethically.
Unlike traditional chat demos, these tests focus on how well AI can finish what it starts, read relevant documents, and stay honest under pressure. This shifts the conversation from whether AI writes well, to whether it works well in real-world, trust-sensitive scenarios.
The Future of Trust in AI
With scores ranging from 73 to 95 in the Crucible League — where a score of just 26 is a do-nothing baseline — these results show that AI models can be more than just capable storytellers. They can be trustworthy partners if tested and reinforced early, in controlled environments.
For decision-makers, the message is clear: don’t wait for a breach to discover whether your AI can be trusted. Test it first, in simulated crises, to ensure it upholds the values and integrity your organization needs.
The real takeaway from this experiment is that AI can be trained and tested to resist manipulation before deployment. By simulating crises—social engineering, document misuse, ethical slips—companies can evaluate and reinforce their AI’s integrity. The results show all models refused manipulative requests, with some even closing deals at full price by reading deeper into company files. This proactive approach to AI trustworthiness offers a new layer of security, ensuring AI acts ethically before it ever joins live systems.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
