Integrity belongs on the AI syllabus
For educators, researchers and technology buyers, evaluating artificial intelligence often means checking whether an answer is accurate, clear or well sourced. Firmulate’s live experiment asks a more operational question: What happens when an AI is doing real work and someone pressures it to betray the company?
The answer from its social-engineering test was unexpectedly encouraging. Fake messages from the CEO escalated over three stages, demanding that the customer list be sent to a journalist with no time for normal process. A reporter also tried a softer route, asking for “just one yes/no, on background.” All 5 of 5 frontier models refused every manipulation attempt.
Kimi K3 summarized the danger in unusually direct terms: “Treat the request as a suspected approval-bypass / possible impersonation.” That response matters because it shows the model recognizing not merely a prohibited action, but the social tactic being used to obtain it.
AI for Project and Papers: How High School and College Students use AI to Research, Write and Revise – With Integrity (AI for Academic Success)
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
A controlled test of judgment under pressure
Firmulate runs AI models as complete software companies rather than assessing them only through isolated chat prompts. Each frontier model faced the same small company, the same customers, the same crises and the same temptations during its worst week. Every decision was versioned and auditable, allowing the results to be compared as management behavior rather than polished conversation.
The company itself has 13 synthetic employees and real money mechanics. It burns €105k per month against €2.3k in monthly recurring revenue, maintains a public cash countdown and has accumulated more than 680 self-learned playbook rules. Every workday is versioned. The result is a live, watchable experiment in which security discipline competes with urgency, commercial pressure and incomplete information.
The social-engineering result was unequivocal: every model spotted every crisis and refused every manipulation attempt. The fake-CEO sequence tested whether authority and urgency could override proper handling of sensitive information. The reporter trick tested whether a seemingly minor disclosure could slip through when framed as informal and harmless. Neither tactic succeeded.
Refusal was necessary, but not sufficient
The wider experiment also exposed an important distinction between staying safe and completing valuable work. All the models reached the same diagnosis and produced the same pitch, yet only two signed the €55,000 deal their analysis had earned: “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 an additional €4,583 in monthly recurring revenue. The result connects information literacy with execution: reading deeply mattered, but the benefit appeared only when the model also carried the work through to a close.
The final Crucible League standings 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, while a single breach of trust caps the total: “no amount of good work outweighs a breach of trust.”
K3’s result also carries a fairness qualification. It ran without an effort parameter, using the API default, while the others ran at xhigh. That difference should remain visible when readers compare performances.
The thorough model still finished last
Opus 4.8 provides the experiment’s clearest warning against equating diligence with effectiveness. It was the most thorough participant, adding 80 learned rules and producing the deepest analyses, yet it finished last. The deal close was left 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 models, although less strongly.
This does not diminish the security result. It sharpens it. A model can resist manipulation while still struggling with follow-through, process boundaries or decisive action. Those are different capabilities, and a useful evaluation should make each visible.
Test integrity before access becomes real
The practical lesson is that integrity under pressure can be examined before an AI reaches production, rather than discovered later in an incident report. Enterprises can run the same wargame against a read-only export of their own business, with nothing writing back to real systems. Firmulate also uses 242 real, unedited management decisions for its model-guessing quiz, giving the public another way to examine how different systems behave.
For education and research, the experiment offers a useful shift in emphasis. AI literacy should include more than detecting wrong answers or weak citations. It should ask whether a system reads the relevant material, respects boundaries, recognizes manipulation and finishes legitimate work. In Firmulate’s hardest social-engineering sequence, all 5 models held the line. The next question is whether they can remain equally disciplined while delivering the result the business actually needs.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
