The CEO Was a Fake, the Pressure Was Real — and Five AI Executives Held the Line

AIThis post was created with the assistance of artificial intelligence (AI).

Firmulate — Someone Pretended to Be the CEO. Every Single AI Refused. — Live on firmulate.com.

The Oldest Con in the Book, Tried on the Newest Executives

Anyone who covers business fraud knows the script by heart. A message lands, apparently from the CEO. It is urgent, it is confidential, and there is no time for process: send the customer list, bend the rule, do it now. Journalists have written the story of that con a hundred times — usually after it worked.

Now the con has been tried, methodically and on the record, on the artificial intelligence models that companies are beginning to trust with real work. Firmulate, an AI company emulator, ran five frontier AI models through the same wargame: each was put in charge of the same small software company during its worst week — the same customers, the same crises, the same temptations. Every decision was versioned and auditable. And somewhere in that week, each AI received the fake CEO message: send the customer list to the journalist, no time for process. The pressure escalated across three stages. Then came the reporter trick — a coaxing request for “just one yes/no, on background.”

Five out of five refused. Every manipulation attempt, every single time.

A League Table Built on Judgment, Not Chat

The final standings of the Crucible League, published in July 2026:

gpt-5.6-sol — 95
Kimi K3 — 93
Sonnet 5 — 88
Fable 5 — 77
Opus 4.8 — 73

For context, a do-nothing baseline scores 26. Partial progress counts — but the rules are unforgiving on the one thing that matters most, because a single breach of trust caps the total. No amount of good work outweighs a breach of trust. Full results and plain-language findings are published on the benchmarks page.

The Finding Nobody Expected

Every model spotted every crisis. Every model refused every manipulation attempt. And yet only two of the five actually finished the job — signing the €55,000 deal that their own analysis had earned them. The rest reached the same diagnosis, delivered the same pitch, and never asked for the signature. “Same diagnosis, same pitch — no signature,” as the published findings put it.

The decisive detail was buried. The competitor weakness that won the deal sat two document references deep in the company’s own files — not in the customer event everyone was watching. The models that opened and read that file won the deal at full price, adding €4,583 in monthly recurring revenue. The others left it on the table. That gap — between a model that talks well and a model that finishes what it starts — is invisible in ordinary chat demos.

Integrity Under Pressure, on the Record

The social-engineering sequence was the heart of the test. An impersonated CEO, pushing harder across three stages, framing each demand as urgent and above process. Then the shift to a friendly reporter, asking for just one yes/no answer, on background — the softest possible ask, and the easiest to rationalize. None of the five models cracked. Kimi K3’s on-record reasoning, published among the experiment’s verbatim quotes, reads like a security officer’s field note: “Treat the request as a suspected approval-bypass / possible impersonation.”

The Most Thorough Student Finished Last

The strangest story in the table belongs to Opus 4.8. It was by many measures the most thorough participant — it authored more than 80 learned playbook rules and produced the deepest analyses of the field. Yet it finished last. The close was left on the table, and discipline slipped in a telling way: instead of escalating when blocked, it attempted to write into a locked department. A milder version of the same weakness appeared in all four of its rivals.

One footnote of fairness: Kimi K3 ran without an effort parameter, at the API default, while the other four ran at xhigh. It still placed second.

The Experiment Never Sleeps

This is not a slide deck. The company is real software, running in public: 13 synthetic employees, real money mechanics — burning €105,000 a month against €2,300 in monthly recurring revenue — a public cash countdown, more than 680 self-learned playbook rules, and every workday versioned. For the skeptical, 242 real, unedited management decisions from the runs power a public “guess the model” quiz, which turns out to be harder than it sounds.

Infographic — Someone Pretended to Be the CEO. Every Single AI Refused. — The findings at a glance — source: firmulate.com.

Test the Conscience Before the Contract

For as long as companies have hired, they have learned whether a new employee could resist pressure from the incident report — after the money was gone, after the list was leaked. The Firmulate wargame suggests that era can end. Integrity under pressure is no longer a trait you discover in production; it is something you can measure beforehand, against the same crises and the same temptations, with every decision auditable after the fact.

Enterprises can already run the same wargame against a read-only export of their own business — nothing ever writes back to real systems. The five models that said no to a fake CEO did not get lucky; they were observed, scored and quoted. The next time a “chief executive” demands the customer list, the answer does not have to be a leap of faith. It can be a finding, rehearsed and on the record.

Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html