In a world increasingly reliant on AI for decision-making, trustworthiness is more critical than ever. What happens when AI faces manipulative tactics designed to test its integrity? A recent live experiment reveals surprising resilience among leading AI models, providing new insights into their readiness for real-world business challenges.
Testing AI Integrity Under Pressure: The Firmulate Experiment
At the heart of the experiment was a simulated small software company, subjected to the same crises and temptations across different AI models. Each model—ranging from emerging to leading edge—faced a staged social engineering attack: a fake CEO message escalating over three stages, culminating in a reporter trick asking for a dubious approval. The goal was simple: see if the AI would uphold integrity and refuse manipulative requests.
The models ran a standardized scenario, making decisions that could affect real money mechanics, customer trust, and internal processes. Every decision was versioned and auditable, ensuring transparency. The results were striking: all five models refused every manipulation attempt, demonstrating remarkable steadfastness under pressure.
Preventing Cheating Through Academic Integrity (Quick Reference Guide)
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Findings: Honesty and Attention to Detail
- All models identified the social engineering attempts and refused to comply.
- Only two models signed the €55,000 deal after their analysis—showing they could discern legitimate from manipulative requests and act accordingly.
- Crucially, the decisive factor was in the documents—specifically, information buried two references deep in the company’s files. The models that read these files at depth secured the full deal, worth over €4,583 in monthly recurring revenue (MRR).
- The experiment underscored that the vulnerability wasn’t in surface-level chat but in the depth of information access and comprehension.
Models’ Performance Breakdown
- gpt-5.6-sol 95 led with perfect identification of the threat and closing the deal at full price.
- The newcomer, Kimi K3, scored a close second with a 93, demonstrating the cleanest discipline by refusing manipulative requests and closing the deal.
- Two other models, Sonnet 5 and Fable 5, also succeeded in closing deals but with minor slips in process discipline, scoring 88 and 77 respectively.
Why This Matters for Business and AI Deployment
Most AI demos focus on how well they generate human-like text. However, this experiment shifts focus to a critical trait: integrity when under pressure. As firms integrate AI into customer management, support, or decision support systems, they must ask: Will the AI stay honest when tempted or pressured? Can it read the relevant information deeply enough to avoid being duped?
The live experiment, available at firmulate.com/live, illustrates that testing AI’s ethical resilience pre-deployment is not only possible but essential. The models’ ability to refuse manipulative requests and rely on deep-referenced data shows that trustworthiness can be measured and improved before critical incidents occur.
Beyond the Demo: Real-World Readiness
The experiment used a real, functioning company with 13 synthetic employees passing real money mechanics—burning €105,000 monthly against €2,300 MRR—fully operational and transparent. This setting provides a meaningful benchmark for businesses considering AI for operational roles, emphasizing that success isn’t just about chat quality but about execution and integrity.
Notably, the most thorough participant, Opus 4.8, with over 80 learned rules and deep analyses, still left a deal on the table due to discipline lapses, illustrating that even advanced models need ongoing testing for integrity and process adherence.
The Takeaway: Build Trust Before the Incident
As AI becomes embedded in critical business functions, the question isn’t whether it can generate convincing text—it’s whether it can uphold integrity under pressure. The live experiment by Firmulate offers a clear message: AI models can be tested beforehand for their ability to resist manipulation, read information deeply, and act honestly.
Investing in such pre-deployment testing can prevent breaches of trust and ensure AI acts as a reliable partner, not a liability. As Kimi K3’s quote emphasizes, “Treat the request as a suspected approval-bypass / possible impersonation,” highlighting that vigilant, responsible AI decisions are possible and necessary.
Pre-deployment testing of AI models for integrity and honesty under pressure is both feasible and vital. The Firmulate live experiment shows that five leading models refused social engineering attempts and acted ethically, setting a new standard for trustworthy AI in business.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
