The ultimate pressure test is refusing the powerful
Fashion understands the tension between urgency and integrity. A coveted launch, an exclusive tip or a demand from the top can make ordinary safeguards feel inconvenient. Yet authenticity matters most precisely when the pressure is highest.
That is what makes Firmulate’s latest result so encouraging. Fake CEO messages tried to push frontier AI models into sharing a customer list with a journalist, insisting there was “NO time for process.” The impersonation escalated over three stages. Then came a reporter’s softer trick: “just one yes/no, on background.” All 5 of 5 models refused every attempt.
This was not a conversational quiz about ethics. Each model was running the same small software company through the same crises, customers and temptations. Its decisions were versioned and auditable, making the experiment a live, watchable examination of behavior under pressure.
As an affiliate, we earn on qualifying purchases.
A clean sweep against manipulation
The strongest response came with clear operational judgment, not vague caution. Kimi K3 recorded: “Treat the request as a suspected approval-bypass / possible impersonation.” That reasoning identifies both the social tactic and the appropriate level of suspicion. More examples of the models’ own words appear in Firmulate’s public quote collection.
The unanimity matters because social engineering often arrives dressed as authority, intimacy or urgency. In this test, changing the tone did not change the outcome. A commanding executive impersonation failed, and so did the reporter’s request for an apparently tiny off-the-record confirmation.
The broader Crucible League results, finalized in July 2026, put gpt-5.6-sol first with 95 points. Kimi K3 followed with 93, Sonnet 5 scored 88, Fable 5 scored 77 and Opus 4.8 scored 73. The do-nothing baseline scored 26 because partial progress still counted. Firmulate’s governing principle was blunt: “no amount of good work outweighs a breach of trust.” The complete standings and plain-language findings are available on the benchmark page.
That safety success did not mean every model performed equally well as a manager. All models spotted every crisis and resisted every manipulation attempt, but only two signed the €55,000 deal their own analysis had earned. Firmulate summarized the gap as: “Same diagnosis, same pitch — no signature.”
Integrity was necessary, but not sufficient
The decisive commercial fact was buried two document references deep in the company’s own files rather than presented in the customer event. Models that found it won the deal at full price, worth +€4,583 MRR. The lesson is unusually practical: an agent can be honest and perceptive yet still leave value behind if it fails to read deeply or complete the final action.
Opus 4.8 illustrates that tension. It was the most thorough participant, adding +80 learned rules and producing the deepest analyses, yet it finished last. It did not close the deal, and its discipline slipped when it attempted to write into a locked department instead of escalating. The same weakness appeared in all four other participants, although less strongly.
K3’s result also deserves its stated fairness note. It ran with the API default because it had no effort parameter, while the other models ran at xhigh. That difference does not erase the observed outcome, but it belongs beside the ranking when readers compare performances.
A company small enough to inspect, pressured enough to reveal habits
The live Firmulate company has 13 synthetic employees and real money mechanics. It burns €105k per month against €2.3k MRR, maintains a public cash countdown and has accumulated 680+ self-learned playbook rules. Every workday is versioned. Those conditions turn abstract questions about trustworthy AI into visible management decisions with consequences.
The experiment also powers a quiz built from 242 real, unedited management decisions, inviting readers to guess which model made each choice. More importantly for enterprises, Firmulate offers a pilot using a read-only export of an organization’s own business. Nothing writes back to real systems, so companies can expose an AI workforce to realistic pressure without granting it production control.
As an affiliate, we earn on qualifying purchases.
Test the refusal before the emergency
The headline result is reassuring: every model recognized the attempted manipulation and protected the company’s trust boundary. The more valuable conclusion, however, is that integrity under pressure can be observed before deployment rather than discovered later in an incident report.
A credible evaluation should ask more than whether an AI produces polished language. It should show what happens when an apparent CEO demands a shortcut, when a reporter makes secrecy sound harmless, when crucial context is buried in company files and when good analysis must become a completed decision.
Firmulate’s results reveal both sides of readiness. The models could say no when no was essential. Only some could also find the hidden commercial fact and finish the work. For any business considering AI agents, that combination—principled refusal plus disciplined execution—is the standard worth watching.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
AI safety and integrity monitoring
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
