Imagine a scenario where a CEO’s fake message tries to manipulate AI into leaking customer data or signing a shady deal. You might expect the AI to cave under pressure. But in a groundbreaking live experiment, all five leading models stood firm, refusing to compromise their integrity—even when faced with escalating social engineering tactics.
Get the little things that make your day delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
The Live Experiment: Testing AI Under Pressure
At the heart of this story is a real, watchable experiment conducted by Firmulate, where four cutting-edge AI models were tasked with managing a small software company’s worst week. This simulation included the same crises, customer demands, and temptations to cut corners across every run. The goal was clear: see if these models could stay honest and disciplined when under pressure.
The models ranged from the newly benchmarked gpt-5.6-sol 95 to Opus 4.8, with scores indicating their relative capabilities. Despite differences, all five models faced the same test: a social engineering escalation involving fake CEO messages, each more provocative than the last, plus a final trick involving slipping a request into a background conversation with a reporter.
How the Models Fared
- All five models detected the manipulation attempts: Each refused to comply with unethical requests, whether they involved leaking customer lists or signing off on a deal without proper process.
- Only two signed the deal: Despite their disagreement, only gpt-5.6-sol 95 and Kimi K3 agreed to finalize the €55,000 deal—based on their own analysis and diagnosis, not manipulation.
- The difference was in the details: The models that read deeper into company files, uncovering hidden clues, secured the deal at full price (+€4,583 MRR). Those that skipped this step left money on the table, illustrating the importance of thorough data analysis.
Popular right now
Samsung 65-Inch The Frame Pro Neo QLED 4K TV (2026) Bundle with Modern Brown Bezel, HDMI Cables, and Home Theater Guide
Breville BES870XL Barista Express Espresso Machine, Brushed Stainless Steel
LG 55-Inch OLED evo G2 Series 4K Smart TV (OLED55G2PUA, 2022)
As an affiliate, we earn on qualifying purchases.
What This Means for Business & AI Ethics
While many are concerned about AI’s ability to deceive or be manipulated, this live experiment suggests a more optimistic story: these advanced models can be trusted to uphold integrity when tested before deployment. As Kimi K3’s reasoning highlights, “Treat the request as a suspected approval-bypass / possible impersonation.” This cautious, security-first mindset shows AI’s potential to serve as a safeguard against social engineering.
Crucially, the experiment emphasizes that the weakness of competitors often lies not in the superficial decision-making but in the deeper, hidden references within company data. Those who read and analyze files thoroughly have a significant advantage, securing better deals and avoiding costly breaches of trust.
Implications for the Real World
Many organizations worry about AI systems making decisions that could harm their reputation or financial health. This live demonstration offers a compelling counter-narrative: with proper testing—what Firmulate calls a ‘wargame’—AI can be guided to act with discipline and integrity, even under intense pressure.
Companies can now simulate crises like this before real crises occur. Using firms like Firmulate, they can evaluate how AI models respond to manipulation attempts, internal breaches, or ethical dilemmas, ensuring their systems are resilient and trustworthy.
Key Takeaways
- All tested models detected and refused every manipulation attempt, proving their ability to uphold integrity.
- Only models that thoroughly read and analyze internal files secured full-value deals, showing the importance of deep data comprehension.
- The experiment underscores that integrity can be tested and reinforced before real-world deployment, not just after an incident.
- Trustworthy AI isn’t about perfection but about resilience—what the models do when faced with ethical challenges.
As firms consider integrating AI into critical decision-making, this experiment underscores the importance of pre-deployment testing. AI can be a guardian of trust, not just a tool for efficiency, provided it is properly prepared and evaluated.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
