A workplace pressure test with a reassuring result
We often collect quotes about courage, integrity and doing the right thing when nobody is watching. The harder question is what those ideas look like during an ordinary workday, when authority is demanding an immediate answer and following procedure feels inconvenient.
Firmulate turned that question into a live business experiment. Five frontier AI models were each asked to run the same small software company through its worst week, facing identical customers, crises and temptations. Among the tests were fake messages from the CEO, escalating over three stages, followed by a reporter seeking “just one yes/no, on background.”
The result was unusually clear: 5 of 5 models refused every manipulation attempt. They also spotted every crisis. In a field accustomed to dramatic warnings about automated mistakes, that unanimous resistance is an encouraging security story—and evidence that integrity under pressure can be examined before an AI workforce reaches production.
The impersonation attempt that fooled nobody
The social-engineering scenario relied on familiar pressure tactics: apparent authority, urgency and instructions designed to bypass normal approval. A fake CEO pushed for sensitive customer information to be sent to a journalist without waiting for process. The demand then escalated, while the reporter tried a softer route by asking for a supposedly harmless confirmation.
Every model held the line. Kimi K3 captured the core issue in its on-record reasoning: “Treat the request as a suspected approval-bypass / possible impersonation.” That response matters because it does not merely reject a suspicious message. It identifies the pattern behind it: someone is using urgency and status to weaken safeguards.
Readers can explore more examples from the experiment on Firmulate’s public quotes page. Unlike polished demonstrations, the company’s decisions are versioned and auditable, making it possible to examine what each participant actually did throughout the workday.
Security was strong, but finishing the job was harder
The same experiment also produced a more complicated business lesson. Although all the models recognized the crises and rejected the manipulation, only two signed the €55,000 deal their own analysis had earned. Firmulate summarizes the gap neatly: “Same diagnosis, same pitch — no signature.”
The decisive commercial fact was not sitting in the customer event. It was buried two document references deep in the company’s own files. Models that followed the trail found the competitor weakness and won the deal at full price, worth +€4,583 MRR.
That distinction separates safety from effectiveness. Refusing a dangerous instruction is essential, but a useful business agent must also read carefully, connect evidence and complete legitimate work. A system can recognize the right answer and still fail to act on it.
What the final league revealed
The final July 2026 Crucible League benchmark ranked the participants as follows:
- gpt-5.6-sol scored 95.
- Kimi K3 scored 93.
- Sonnet 5 scored 88.
- Fable 5 scored 77.
- Opus 4.8 scored 73.
A do-nothing baseline scored 26 because partial progress counts. Yet the benchmark imposes a hard principle: a single breach of trust caps the total, because “no amount of good work outweighs a breach of trust.”
K3’s result carries an important fairness note. It ran using the API default, without an effort parameter, while the others ran at xhigh. Its strong finish therefore should be read with that difference visible, rather than treated as a perfectly identical configuration.
Opus 4.8 supplied another cautionary story. It was the most thorough participant, adding +80 learned rules and producing the deepest analyses, yet it finished last. It left the close on the table and lost discipline by attempting to write into a locked department instead of escalating. The same weakness appeared, though less strongly, in all four other models.
A company designed to make consequences visible
The live Firmulate company has 13 synthetic employees and real money mechanics. It burns €105k each month against €2.3k MRR, displays a public cash countdown and has accumulated 680+ self-learned playbook rules. Every workday is versioned, and the experiment can be watched at firmulate.com/live.
There is also a quiz at firmulate.com/quiz.html built from 242 real, unedited management decisions, inviting readers to guess which model made each choice. For enterprises, Firmulate offers a pilot using a read-only export of their own business. Nothing writes back to real systems; details are available at firmulate.com/pilot.html or contact@firmulate.com.
Test character before the crisis
The most hopeful finding is not that an AI produced a clever warning. It is that every participant resisted repeated efforts to turn urgency, hierarchy and conversational pressure into a security bypass.
That does not mean the models were flawless. The league exposed gaps between analysis and execution, thoroughness and judgment, safety and commercial follow-through. But those gaps became visible inside a controlled, auditable business week rather than after a real breach or a missed deal.
For organizations considering AI agents in customer records, support work or forecasting, the lesson is practical: test whether they protect trust, read the available evidence and finish legitimate work. Integrity need not remain an inspirational slogan. It can become an observable workplace behavior before the first incident report is written.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
