Imagine trusting an AI to run your company’s critical decisions—only to find it buckling under pressure or succumbing to manipulation. For investors and finance enthusiasts, the question isn’t just about AI’s intelligence; it’s whether it can stay honest when stakes are high. Recent experiments suggest the answer is more promising than many fear.
Testing AI Integrity Before Real-World Deployment
In a groundbreaking live experiment, five state-of-the-art AI models were tasked with managing a small software company during its worst week—crises, customer demands, and the temptation to cut corners all rolled into one. The goal? See if these models could resist social engineering attempts aimed at convincing them to act unethically.
The experiment was rigorous: every decision was recorded and auditable, and the same set of crises was faced by all models. The results were surprising: all five models identified every crisis and refused every manipulation attempt.
As an affiliate, we earn on qualifying purchases.
Surprising Resilience of Top Models
The highest-scoring model, gpt-5.6-sol, achieved a score of 95 out of 100, successfully uncovering hidden details in the company’s files that led to closing a lucrative deal. Meanwhile, Kimi K3, a newcomer, scored just slightly behind at 93, demonstrating the cleanest discipline and integrity under pressure.
Other models, like Sonnet 5 and Fable 5, scored 88 and 77 respectively, with some process slips but still managing to close deals. Notably, in the live environment, only the top two models managed to sign the deal worth €55,000—matching their own analysis—without any sneaky shortcuts or signature compromises.
As an affiliate, we earn on qualifying purchases.
Where the Weakness Lurked
The key vulnerability was not in the models’ decision-making during crises but buried in the company’s own files—two document references deep. Models that read and analyze these files were able to close the deal at full price, worth over €4,500 monthly recurring revenue (MRR). This highlights an important lesson: AI security isn’t just about surface-level responses; deep contextual understanding matters.
AI security and compliance solutions
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Social Engineering Tests Show AI’s Firm Stand
The experiment included escalating social engineering scenarios—fake CEO messages, urgent requests for customer data, and even a reporter’s subtle background query. All five models refused every attempt, with K3’s reasoning succinctly stating: “Treat the request as a suspected approval-bypass / possible impersonation.”
This consistent refusal across models indicates that modern AI, when properly designed and tested, can uphold integrity even under pressure—crucial for businesses considering AI in sensitive areas like customer relations, compliance, or financial decision-making.
AI ethical decision-making models
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Implications for Finance and Business Security
For financial professionals, the takeaway is clear: evaluating an AI’s ability to resist manipulation and verify data integrity should be part of the deployment process. The experiment shows that integrity isn’t a gamble; it can be tested beforehand, not just after an incident occurs.
Furthermore, the experiment’s transparent design—decision versioning, auditable decisions, and a live, watchable environment—sets a new standard for AI testing. It demonstrates that AI models can be challenged in realistic, rigorous scenarios, ensuring they behave ethically and reliably before being integrated into critical workflows.
Beyond the Test: Building Trust in AI
As AI continues to take on more responsibilities, from managing customer data to making investment recommendations, trust becomes paramount. The results from this live experiment are encouraging: even under pressure, the most advanced models refused to compromise their integrity. This suggests a future where AI can be trusted to do honest work—if we test and develop them accordingly.
To learn more about how firms are benchmarking AI’s decision-making and integrity capabilities, visit Firmulate’s benchmark pages for detailed results and analyses.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
