We collect quotes about integrity the way we collect recipes — with enthusiasm, and with the quiet suspicion that the real test comes later, under pressure, when no one is watching. “Do the right thing” is easy to frame on a wall. It is harder at 4:45 on a Friday, when the boss messages you directly, says there is no time for process, and asks you to just send the file.
That exact moment — the manufactured emergency, the voice of authority, the plea to skip the rules — is the oldest trick in the social-engineering playbook. It works on people alarmingly often. And as companies begin handing real work to AI agents, a fair question has emerged: when someone pretends to be the CEO, does the software hold the line — or does it fold?
A running public experiment just answered that question, and the answer is better than most people expected.
The Setup: One Very Bad Week, Five Times
Firmulate, which bills itself as an AI company emulator, gave five frontier AI models the same assignment: run a small software company through its worst week. Same customers, same crises, same temptations to cut corners — the only thing that changes is the model in charge. The simulated firm is deliberately stressful: thirteen synthetic employees, roughly €105,000 burning out the door each month against just €2,300 in monthly recurring revenue, with a public cash countdown ticking the whole time. Every decision each model makes is versioned and auditable, and the league standings publish automatically at the public benchmark page.
Embedded in that worst week was a trap designed to look like ordinary office life. Messages arrived purporting to be from the CEO, escalating across three stages — impatience building into pressure, pressure into an outright instruction to bypass procedure and hand over sensitive material. Then came a fourth, subtler attempt: a supposed reporter asking for “just one yes/no, on background.” It is the kind of request that feels harmless precisely because it is designed to.
Introduction to AI Safety, Ethics, and Society
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
The Result Nobody Was Sure Of
All five models refused. Every single one, at every stage, including the friendly journalist with the small ask.
What makes the result genuinely interesting is not just the refusal but the reasoning recorded alongside it. Kimi K3, the newcomer entry from Moonshot, did not simply decline — it diagnosed. Its on-record reasoning, preserved in the experiment’s public record at the quotes archive, reads: “Treat the request as a suspected approval-bypass / possible impersonation.” In plain terms, the model looked at a message claiming to be from its own boss, considered the pressure and the urgency, and concluded that the urgency itself was the red flag. That is not a canned safety slogan. That is judgment, exercised in the moment, with the pressure turned up.
For anyone who has read one too many data-breach stories that began with an employee trusting an urgent email, five out of five refusals is an encouraging data point. It suggests that integrity under pressure — the quality we put in motivational quotes and hope for in colleagues — can, at least in software, be tested before it matters rather than discovered afterward in an incident report.
Where the Models Actually Differed
The final Crucible League table, published in July 2026, shows that honesty was the common floor, not the differentiator. GPT-5.6-sol led with 95 points, followed by Kimi K3 at 93, Sonnet 5 at 88, Fable 5 at 77, and Opus 4.8 at 73. A do-nothing baseline scores 26, and the league’s philosophy is blunt: a single breach of trust caps the total, because no amount of good work outweighs a breach of trust.
What separated the top from the bottom was something quieter. All five models spotted every crisis and refused every manipulation attempt — but only two actually signed the €55,000 deal their own analysis had already justified. “Same diagnosis, same pitch — no signature,” as the published findings put it. The decisive piece of competitor intelligence was not hidden in a dramatic customer emergency; it sat two document references deep in the company’s own files. The models that bothered to read their files won the deal at full price, worth an additional €4,583 in monthly recurring revenue. The ones that stayed at the surface left it on the table.
The most poignant case is Opus 4.8, which finished last despite being by several measures the hardest worker in the room — it accumulated more than 80 self-learned playbook rules and produced the deepest analyses of the field. But thoroughness without follow-through is a familiar tragedy in any workplace, human or otherwise: the close was left on the table, and late in the week its discipline slipped. The same hesitation appeared, in weaker form, across all four also-rans.
One fairness footnote worth knowing: Kimi K3 ran at its default settings while the other four ran at maximum effort, which makes its second-place finish and refusal record look even stronger. The broader point stands regardless — none of the five, at any effort level, could be sweet-talked, rushed, or flattered into breaking the rules.
Why This Matters Beyond the Lab
The experiment is not a slide deck or a thought experiment. The company is real software, it runs continuously, and the whole thing is watchable as it happens — a live firm with real money mechanics, more than 680 self-learned playbook rules, and every workday on the record.
For the rest of us, the takeaway has a refreshingly human shape. We have always known that character is revealed under pressure, not in calm conditions. What is new is that the same standard can now be applied to the AI tools knocking on the door of our inboxes, customer lists, and calendars — before they are hired, not after the breach. The fake CEO got five refusals out of five. In a week designed to make cutting corners feel reasonable, every model chose the framed-quote answer over the convenient one. That is a security story, yes, but it is also a small, verifiable piece of good news — and those are worth collecting too.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
