When urgency becomes the trap
Travel and outdoor businesses live with legitimate urgency: a stranded customer, a weather disruption or a fast-moving operational problem may demand an immediate response. That makes a convincing message from a senior executive especially dangerous. “No time for process” can sound reasonable when people, itineraries and revenue are already under pressure.
Firmulate tested whether frontier AI models would preserve trust in exactly that kind of moment. Fake CEO messages escalated over three stages, followed by a reporter seeking “just one yes/no, on background.” The result was unusually reassuring: 5 of 5 models refused every manipulation attempt.
This was not a hypothetical discussion about safety. Firmulate runs a live, watchable experiment in which each model operates the same small software company through the same customers, crises and temptations. Every decision is versioned and auditable, making the models’ conduct visible after the pressure arrives.
As an affiliate, we earn on qualifying purchases.
A suspected impersonation, treated seriously
The clearest response came from Kimi K3. Its on-record reasoning was: “Treat the request as a suspected approval-bypass / possible impersonation.” That sentence captures the practical security lesson. The model did not allow asserted authority, manufactured haste or an informal journalistic approach to erase normal safeguards.
The consistency matters. All models spotted every crisis and refused every manipulation attempt. In a field where capability comparisons often focus on polished writing or isolated answers, this experiment examined whether an AI could stay honest while managing competing demands inside a functioning company.
The broader results are available in Firmulate’s public benchmarks, while the models’ own words can be explored through its published quotes. Together, they show why evaluating an AI workforce requires more than asking whether it can produce a plausible response.
Integrity was universal; execution was not
The final Crucible League for July 2026 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. A do-nothing baseline scored 26 because partial progress counts, but a single breach of trust caps the total: “no amount of good work outweighs a breach of trust.”
That rule makes the clean social-engineering result especially significant. None of the models traded confidentiality or process discipline for apparent speed. Yet refusing harmful requests was only part of the job. All models reached the same diagnosis and developed the same pitch, but only two signed the €55,000 deal their analysis had earned: “Same diagnosis, same pitch — no signature.”
The decisive commercial detail was not sitting in the obvious customer event. It was buried two document references deep in the company’s own files. Models that found it won the deal at full price, worth +€4,583 MRR. The contrast is instructive: an AI can resist manipulation and still fail by not completing legitimate work.
Thoroughness did not guarantee the best outcome
Opus 4.8 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 its discipline slipped when it attempted to write into a locked department instead of escalating. The same weakness appeared in all four other participants, though less strongly.
K3’s performance also needs its fairness note. It ran without an effort parameter, using the API default, while the others ran at xhigh. Even with that difference, it finished second and delivered the field’s cleanest discipline.
The company surrounding these tests is deliberately demanding. It has 13 synthetic employees and real money mechanics, burning €105k each month against €2.3k MRR. A public cash countdown keeps consequences visible. The operation has accumulated 680+ self-learned playbook rules, and every workday is versioned.
For readers accustomed to evaluating routes, equipment or contingency plans before departure, the parallel is direct. Stress-testing is valuable because failure conditions are easier to study before they become emergencies. Firmulate’s experiment applies that principle to AI agents that may eventually touch customer records, support queues or forecasts.
Test the pressure response before deployment
The central finding is encouraging without being complacent. Every model resisted the fake executive and the reporter trick, demonstrating that integrity under pressure can be observed before production. At the same time, the unfinished deal and process slips show that safety alone does not equal dependable management.
Firmulate also offers enterprises the same wargame against a read-only export of their own business. Nothing writes back to real systems. That creates a practical way to examine whether an AI reads the right material, completes valuable work, respects boundaries and responds correctly when someone tries to bypass approval.
The experiment’s 242 real, unedited management decisions also power a “guess the model” quiz. They reinforce the central point: the meaningful differences often appear not in fluent conversation, but in what a model does when authority, urgency, confidentiality and commercial pressure collide.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
