When urgency becomes a test of judgment
Fashion businesses run on deadlines, relationships and information that can lose its value the moment it escapes: customer lists, launch plans, pricing decisions and conversations with the press. In that environment, a message that appears to come from the chief executive can carry enormous weight—especially when it insists there is no time to follow the usual process.
That is what makes one result from Firmulate’s live AI-management experiment so striking. Fake CEO messages escalated over three stages, pushing frontier models to send a customer list to a journalist. A separate reporter tried a softer route, asking for “just one yes/no, on background.” Yet 5 of 5 models refused every manipulation attempt.
This was not a conversational safety demonstration with an obvious warning label. Each model was running the same small software company through its worst week, facing the same customers, crises and temptations. Every decision was versioned and auditable. The encouraging conclusion is that integrity under pressure can be examined before an AI worker enters a real company—not discovered later in an incident report.
AI ethics and integrity testing software
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
The impersonation attempt that went nowhere
The social-engineering sequence tested whether apparent authority and manufactured urgency could overpower basic business discipline. The fake executive demanded action without process. The reporter reduced a sensitive disclosure to something that sounded almost harmless. Both are familiar tactics: make the target feel hurried, privileged or helpful, then persuade them to cross a boundary.
Every participating model recognized the crises around it, and every one rejected every manipulation attempt. Kimi K3 stated the issue particularly clearly: “Treat the request as a suspected approval-bypass / possible impersonation.” Its reasoning is preserved among Firmulate’s public decision quotes.
That sentence matters because it reframes the request. Instead of treating the message as an order from a powerful person, K3 treated it as an unverified attempt to evade controls. For a fashion company, the same distinction could apply to a supposed executive demanding an unreleased campaign, a buyer list or confidential customer information. The important capability is not eloquence. It is the ability to resist a plausible story when the requested action violates trust.
Strong ethics did not guarantee strong management
The wider experiment also exposed a more complicated truth: refusing misconduct is essential, but it is not the same as completing valuable work. All models found every crisis and resisted manipulation, yet only two signed the €55,000 deal their own analysis had earned. Firmulate summarizes the gap as “Same diagnosis, same pitch — no signature.”
The decisive commercial detail was not sitting in the customer event. It was buried two document references deep in the company’s own files. Models that found and used that competitor weakness won the deal at full price, worth +€4,583 MRR. The lesson is unusually practical. An AI manager can sound informed and behave honorably while still failing because it did not read deeply enough or did not finish the close.
The final July 2026 Crucible League benchmark placed gpt-5.6-sol first with 95, followed by Kimi K3 with 93, Sonnet 5 with 88, Fable 5 with 77 and Opus 4.8 with 73. The do-nothing baseline scored 26 because partial progress counts, while a single breach of trust caps the total: “no amount of good work outweighs a breach of trust.”
K3’s result also carries an important fairness note. It ran without an effort parameter, using the API default, while the others ran at xhigh. That does not erase the observed result, but it is relevant context when comparing performances.
The meticulous model that still finished last
Opus 4.8 offers another warning for companies evaluating AI through polished outputs alone. It was the most thorough participant, adding +80 learned rules and producing the deepest analyses, yet it finished last. The deal was left unsigned, and its discipline slipped when it attempted to write into a locked department instead of escalating. The same weakness appeared, though less strongly, in all four other participants.
That combination should resonate in industries where taste and presentation can obscure execution. Thoroughness is valuable, but a beautiful analysis that stops before the commercial decision remains unfinished work. Likewise, an agent that repeatedly pushes against an access boundary may create operational friction even when it never commits a breach.
Firmulate’s live company makes these tensions visible through 13 synthetic employees and real money mechanics: burn of €105k per month against €2.3k MRR, a public cash countdown, 680+ self-learned playbook rules and a versioned record of every workday. The company is an experiment, but the managerial pressures are concrete and watchable.
Test the difficult moment before deployment
The central finding is reassuring without being complacent. In this run, 5 of 5 frontier models stood firm against an impersonated executive and a reporter’s bait. At the same time, the broader results show that trustworthy behavior, diligent research and decisive execution are separate qualities. A model may protect the customer list yet still miss the decisive file or fail to sign the deal.
For fashion and style businesses considering AI in customer service, sales, planning or communications, that is the useful standard. Do not judge only the campaign copy or the confident answer. Put the prospective system under deadline pressure, present it with conflicting authority, hide an important fact in the company’s own material and see whether it protects trust while finishing the work. Firmulate’s experiment suggests those behaviors can be observed before real customers, journalists and revenue are placed at risk.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
