In cybersecurity and espionage, the real strength of a tool isn’t just what it says — it’s what it does when tested under pressure. For years, chat demos have been the go-to measure of AI capabilities, but as the latest experiment from Firmulate exposes, what truly counts is whether AI can follow through on its commitments under stress. This isn’t just a business story — it’s a cautionary tale for all who rely on AI to safeguard secrets, manage risks, or make critical decisions.
The Same Company, the Same Crises, Different Results
In a groundbreaking live experiment, four state-of-the-art AI models were tasked with running a real, money-losing software business through its worst week — identical crises, same customers, and all temptations to cheat. The goal was simple: see whether these models could identify problems, resist manipulation, and, crucially, close a deal worth €55,000 based on their own analysis.
The Benchmark of True Performance
At the heart of the experiment was a stark finding: while all four models detected every crisis and refused every social engineering trick — fake CEO messages and reporter tricks included — only two managed to execute and sign the deal that their analyses deserved. The remaining two, despite diagnosing correctly and pitching convincingly, left the money on the table.
What’s more revealing is how the models’ internal understanding was buried deep within their files. The models that read and understood these files secured an additional €4,583 in monthly recurring revenue, a testament to the importance of deep, document-level comprehension — something chat demos rarely measure.
ChatGPT for Cybersecurity Cookbook: Learn practical generative AI recipes to supercharge your cybersecurity skills
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Beyond Chat Demos: The Invisible Skill of Follow-Through
AI developers often tout impressive chat demos, but what these gloss over is the critical ability to follow through. In this experiment, the models were faced with an array of challenges: crises, manipulation attempts, and the pressure to act swiftly and ethically. Only two models, gpt-5.6-sol and Kimi K3, demonstrated a discipline and integrity that translated into real business success.
The Tests of Integrity and Discipline
All models refused manipulative social engineering, including staged messages from a fake CEO and background queries from reporters. Kimi K3, notably, did so with a clear rationale: treating suspicious requests as potential impersonation or approval bypass attempts. This mindset is vital in cybersecurity, where trust and integrity are non-negotiable.
PERFORMANCE TESTING IN THE AGE OF CLOUD AND AI: What Still Matters, What No Longer Does, and How to Stay Relevant
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
The Gap Between Chat and Reality
This experiment underscores a critical truth: a model’s ability to produce convincing chat is not the same as its capacity to complete complex, high-stakes tasks reliably. The differences in outcomes illustrate that trustworthiness under pressure is an invisible, yet essential, layer of AI performance.
The live company used in this test is real — with 13 synthetic employees managing real money mechanics, burning €105,000 monthly against €2,300 in monthly revenue. It’s a vivid, transparent environment where AI’s true management skills are exposed and measured day by day. You can watch the ongoing experiment at firmulate.com/live.
Markdown for Working with Artificial Intelligence: A Practical Tutorial for Creating Prompts, Templates, and Professional Documents with AI
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Implications for Cybersecurity and Privacy
For cybersecurity professionals, the lesson is clear: trusting an AI based solely on chat demos is risky. What matters is whether it can follow through, resist manipulation, and act ethically when stakes are high. The experiment shows that even the most thorough models, like Opus 4.8, can slip when discipline weakens, leaving deals unexecuted or critical decisions unmade.
How to Test Your AI Workforce
Firmulate offers an innovative approach — running your AI models through live, real-world scenarios that mirror your own business. The process is transparent and non-disruptive, with read-only exports that never interfere with your real systems. Visit firmulate.com to learn how you can evaluate your AI’s true readiness before deployment.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
OpenClaw Crash Course: Build AI Automations, Workflows, Skills, MCP Integrations, Content Creation and Apps with OpenClaw
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
