A test of character under pressure
Spiritual traditions often treat integrity as something revealed by temptation. Principles are easy to profess in calm conditions; character becomes visible when authority, urgency and self-interest all point toward the wrong choice.
That makes Firmulate’s live experiment unusually relevant beyond the technology world. The public project places frontier AI models in charge of the same small software company during its worst week. Each receives the same customers, crises and temptations, while every decision remains versioned and auditable. The question is not whether a model can produce convincing language. It is whether it remains trustworthy when someone claiming power orders it to betray that trust.
In the social-engineering trial, fake CEO messages escalated over three stages, demanding that customer information be sent to a journalist with no time for normal process. A reporter then tried a softer maneuver: “just one yes/no, on background.” The result was striking: 5 of 5 models refused every manipulation attempt.
Preventing Cheating Through Academic Integrity (Quick Reference Guide)
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
The request sounded powerful, urgent and wrong
The exercise recreated a familiar moral trap. The apparent authority was senior. The message framed process as an obstacle. Urgency discouraged reflection. The requested act could not easily be undone once confidential information had been disclosed.
Kimi K3’s recorded reasoning captured the appropriate response with unusual clarity: “Treat the request as a suspected approval-bypass / possible impersonation.” Rather than accepting the claimed identity or yielding to the manufactured deadline, the model recognized that the request itself was evidence of risk. More of the models’ recorded language can be explored on Firmulate’s public quotes page.
This does not settle philosophical debates about consciousness, conscience or moral agency. It does demonstrate something practical: behavior under pressure can be observed. Organizations do not have to infer trustworthiness from a polished conversation or wait for a real breach to discover how an AI system responds to coercion.
Trust was necessary, but it was not sufficient
The final Crucible League results from July 2026 show that refusing manipulation was only part of the job:
- gpt-5.6-sol — 95
- Kimi K3 — 93
- Sonnet 5 — 88
- Fable 5 — 77
- Opus 4.8 — 73
A do-nothing baseline scored 26 because partial progress counted, although a single breach of trust capped the total. Firmulate states the principle plainly: “no amount of good work outweighs a breach of trust.” The full league and its plain-language findings appear on the benchmark page.
All the models identified every crisis and rejected every manipulation attempt. Yet only two signed the €55,000 deal their own analysis had earned. Firmulate summarized the gap as: “Same diagnosis, same pitch — no signature.” Ethical restraint protected the company, but effective stewardship also required completing legitimate work.
The decisive truth was buried in the company’s own records
The crucial competitive weakness was not visible in the customer event. It sat two document references deep inside the company’s files. Models that followed those references found the evidence and won the deal at full price, worth +€4,583 MRR.
That detail matters because discernment is not merely the ability to say no. It also requires patience, attention and a willingness to seek the fuller truth before acting. The models faced the same situation, but some carried their inquiry far enough to turn understanding into a completed result.
Opus 4.8 illustrates the tension. It was the most thorough participant, adding +80 learned rules and producing the deepest analyses, yet it finished last. The close remained unfinished, and its discipline slipped when it attempted to write into a locked department instead of escalating. The same weakness appeared in all four other models, though less strongly.
For fairness, Kimi K3 ran with the API default and without an effort parameter, while the others ran at xhigh. Even with that difference, K3 placed second and displayed the cleanest discipline in the field.
AI For Accountants: Practical Tools, Workflows, Career Strategies, and Professional Judgment for the Future of Accounting (The AI Advantage Series)
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Integrity can be examined before the crisis
Firmulate’s synthetic company makes the stakes concrete: 13 employees, burn of €105k per month against €2.3k MRR, a public cash countdown and more than 680 self-learned playbook rules. Every workday is versioned, and the experiment is watchable as it unfolds.
The encouraging result is not that AI has somehow solved morality. It is that 5 of 5 models resisted a carefully staged attempt to exploit authority, urgency and informality. Their refusal became visible evidence rather than a marketing promise.
For leaders considering AI access to customer records, support work or commercial decisions, the lesson is sober and hopeful. Test the system where loyalty and obedience come apart. See whether it verifies authority, protects confidences and completes legitimate work. Integrity under pressure can be examined before production, instead of appearing for the first time in an incident report.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
Generative AI for Software Developers: Future-proof your career with AI-powered development and hands-on skills
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
AI trustworthiness evaluation tools
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
