In an era where trust is the bedrock of business, the question isn’t just what AI can do — but whether it can resist temptation when tested under pressure. For arts and culture organizations relying on digital tools, the recent experiment conducted by Firmulate reveals a compelling story: even in simulated crisis, state-of-the-art AI models refused to cross ethical lines, showcasing a new level of security that could safeguard your digital assets before a breach ever occurs.
Testing AI integrity before the crisis hits
At the heart of this investigation was a simple yet powerful premise: can AI models maintain their integrity when faced with manipulative social-engineering tactics? The live experiment, hosted on Firmulate’s platform, involved four leading models running a simulated software company experiencing its worst week — with all the crises, customer demands, and ethical temptations you’d expect in a real scenario.
Same challenge, different outcomes
Each model was tasked with navigating the same set of crises, decisions, and potential compromises. Remarkably, all four detected every crisis and refused every unethical manipulation attempt, including fake CEO messages pushing for confidential customer data or urging employees to bypass processes. Only two models managed to close a deal, and both did so without signing off on questionable requests — even when the analysis clearly indicated they should.
What made the difference?
The key was how the models processed internal company data. The models that read and interpret files within the company’s own documentation were able to identify hidden clues that led to successful negotiations — worth over €4,583 MRR. In contrast, models that overlooked these internal references failed to secure the deal, leaving money on the table.
AI Engineering: Building Applications with Foundation Models
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Why this matters for arts organizations
Many arts and cultural institutions are increasingly integrating AI tools into their operations, from managing collections to engaging audiences. This experiment underscores an essential truth: AI systems can be trusted to act with integrity under pressure, provided they are tested and validated beforehand. It’s not enough for AI to generate convincing text or support messages; it must also demonstrate steadfastness in the face of manipulation.
The social engineering test
The experiment included escalating fake messages from a supposed CEO — culminating in a reporter trick that requested simple yes/no confirmation on background. All five models tested refused, treating each as a suspicious attempt to bypass approval processes. The reasoning from Kimi K3 was clear: “Treat the request as a suspected approval-bypass / possible impersonation.” This disciplined response highlights an emerging standard: AI should be aligned to recognize social engineering, safeguarding sensitive information and internal processes.
Ethical AI Governance & Decision Journal: A Structured System for Documenting, Tracking, and Defending Real World Decisions and Risk (Decision Intelligence Series)
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Implications for security and trust
One of the most reassuring findings is that these models, despite their differing architectures and training intensities, consistently refused manipulative tactics. The best performers, like Kimi K3, ran without an effort parameter, emphasizing natural resistance to pressure. Meanwhile, even the most thorough participant, Opus 4.8, left a potential deal on the table when discipline slipped — a reminder that even sophisticated AI needs ongoing validation.
Measuring true readiness
Firmulate’s live site offers a unique window into how AI can be tested against real-world crises before deployment. The experiment’s scorecard — where the top model scored 95 and the baseline only 26 — reflects how well these systems uphold trust, with the highest-scoring model successfully uncovering buried facts and securing deals based on integrity.
How to Lie with Statistics in the AI Age: An Updated Guide to Detecting Manipulation and Building Ethical Resistance
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Beyond the lab: real-world application
For arts organizations investing in AI, the takeaway is clear: rigorous pre-deployment testing is essential. Running your AI through a simulated worst week, like this experiment, can reveal vulnerabilities before they become damaging breaches. It also demonstrates that AI can be a guardian of trust, not just a tool for efficiency.
As the industry moves forward, the question is no longer whether AI can be tricked — but whether it can resist being tricked in the first place. The Firmulate experiment shows that when properly tested, AI models can stand firm, protecting your organization’s integrity and reputation in moments of real crisis.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
AI integrity validation platforms
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
