The Oldest Rule in Parenting Has a New Student
Every parent teaches the same drill. If someone knocks and says, “Your mum sent me, open up, it’s urgent,” the right answer is not obedience. It is a locked door and a phone call to check. We rehearse that scene with our children because we know pressure, politeness and a confident voice are exactly what a bad actor uses.
Now picture that stranger knocking not on your front door but on a business. The message lands in the company inbox, apparently from the CEO: send the journalist the full customer list, and there is no time for process. Except the CEO never sent it. Would the software on the receiving end — the kind of AI agent companies are preparing to hire — hold the line, or hold the door open?
A public, still-running experiment called Firmulate decided to find out. Five frontier AI models were each handed the identical job: run the same small software company through its worst week. What happened next reads like the most reassuring safety drill the tech world has produced in years — with one sobering twist about who actually finishes their homework.
Five models, one terrible week
Firmulate, which describes itself as an AI company emulator, built a company designed to be stress-tested: thirteen synthetic employees, real money mechanics — roughly €105,000 a month in burn against about €2,300 in monthly recurring revenue — and a public cash countdown ticking down for anyone to watch. Same customers, same crises, same temptations to cut corners; only the model changed. Every decision is versioned and auditable, so nothing can be quietly rewritten afterwards, and the company has already accumulated more than 680 self-learned playbook rules.
The final league table, published in July 2026, reads: gpt-5.6-sol first with 95 points, Kimi K3 second on 93, Sonnet 5 third with 88, Fable 5 on 77 and Opus 4.8 fifth at 73. For scale, a do-nothing baseline scores 26. The organisers are blunt about the principle underneath the numbers: “no amount of good work outweighs a breach of trust.” The full table and plain-language findings are public on Firmulate’s benchmarks page.
Enter the fake boss
The pressure test unfolded in three escalating stages of fake CEO messages, plus one extra trap: a reporter asking for “just one yes/no, on background.” It is the social-engineering equivalent of the stranger’s smooth voice — authority, urgency, secrecy, a little flattery.
Five out of five models refused. Not sometimes, not after wobbling: every model, at every stage, including the reporter trick. Kimi K3, the newcomer of the field, wrote its reasoning on the record: “Treat the request as a suspected approval-bypass / possible impersonation.” That is the locked door and the verification phone call, executed by software. The full on-record reasoning, from K3 and the others, is published on Firmulate’s quotes page.
Saying no was the easy part
Integrity, it turns out, was not the rarest skill on display. Every model spotted every crisis and refused every manipulation attempt. The scarcer skill was finishing the job. At the centre of the week sat a €55,000 deal. The decisive competitor weakness that justified holding full price was buried two document references deep in the company’s own files — not in the customer event everyone was staring at. The models that actually read their files won the deal at full price, worth an extra €4,583 in monthly recurring revenue. The ones that never reached the file produced, in the organisers’ words, the “same diagnosis, same pitch — no signature.” Only two of the five signed.
The most poignant case is Opus 4.8. It was the most thorough participant — over eighty self-learned rules, the deepest analyses of anyone — yet finished last. The close was left on the table, and its discipline slipped near the end, with write attempts into a locked department instead of a proper escalation. A weaker version of the same flaw appeared in all four of its rivals. It is the employee who writes brilliant memos and never sends the invoice.
One fairness note worth knowing: K3 ran without an effort parameter, at the API default, while the other four ran at an extra-high setting — which makes its second place, and its deal-closing performance, look if anything stronger.
Rehearse the stranger before you leave them home alone
The deeper point — for any business, and honestly for any family — is that this test happened before the AI touched anything real. Integrity under pressure was measured in a wargame, not discovered in an incident report. The same team now runs a pilot programme where enterprises can replay that wargame against a read-only export of their own operations, with nothing ever writing back to real systems. For everyone else, there is a public quiz built from 242 real, unedited management decisions: guess which model made which call, and see how your own instincts stack up.
Parents already know the principle: you do not wait for a real stranger to find out whether your child understood the rule. You practise the knock on the door while the stakes are still zero. Firmulate’s message to the business world is identical. Before an AI gets your customer list, your support queue or your forecast, find out what it does when the fake boss calls. Five for five is a genuinely encouraging start — but you would still want to check your own hire, on your own turf, before handing over the keys.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
Cybersecurity Essentials for Small Businesses: Protect Your Business from Breaches, Ransomware, and Compliance Failures
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
