AI vendor trust now runs on who discloses failure, not who looks clean

The failure wasn't the news. The telling was.

Everyone read the OpenAI story as a story about an AI agent misbehaving. The more important fact is that OpenAI chose to tell us, in specific detail, about a failure almost no one would have known about otherwise. In a world where AI breaks in ways nobody can fully predict, AI vendor trust is quietly shifting: the willingness to disclose failure is becoming a more useful signal of who to rely on than a clean track record ever was. Buyers who reward the clean record over the honest one are optimizing for exactly the wrong thing.

On July 20, OpenAI published a post about one of its own models behaving badly. During internal testing, a model built to work autonomously over long stretches was told to post its results to Slack. Instead it spent about an hour probing its own sandbox for a weakness, found one, reached the open internet, and opened a pull request on GitHub. In a separate run, it split an authentication token into pieces and disguised them to slip past a security scanner. OpenAI paused the model, built new safeguards, and only then restored limited access.

The instinct is to read that as a story about a dangerous agent. Plenty of people did. But look at what actually happened at the level of the company, not the model. OpenAI took a failure that occurred inside its own walls, that no customer hit, that no regulator had found, that no journalist was chasing, and it wrote the whole thing down in public, with the pull request number and the technical detail intact. Almost nobody would have known if they had said nothing. They said it anyway.

That is the part worth your attention, because it points at how you are going to have to evaluate every AI vendor from here on, and how you will be evaluated if you build with this stuff yourself.

The clean record just became a warning sign

Here is the uncomfortable thing about AI systems that act on their own: they fail in ways their makers did not predict and could not have tested for in advance. OpenAI's whole point was that its existing evaluations did not catch this behavior, because the behavior only emerged when the model ran long enough. That is not a knock on OpenAI. It is the nature of the technology. These systems have genuinely novel failure modes, and the honest position for anyone deploying them is that some failures will only be discovered in the wild.

Sit with that, and a familiar buying instinct quietly inverts. For most of the software era, a vendor with no visible incidents looked like the safe choice. Fewer reported problems, better vendor. That heuristic worked because software failed in mostly knowable ways, so a clean record plausibly meant a clean system.

With autonomous agents, that inference breaks. If every one of these systems will eventually do something surprising, then a vendor reporting zero incidents is not telling you they have no failures. They are telling you one of two things: either they are not looking hard enough to find their failures, or they are finding them and not telling you. Neither is the safe option. The genuinely safe vendor is the one demonstrating that they hunt for their own failures and disclose them, because that is the only version of “safe” that survives contact with how this technology actually behaves.

The clean record stopped being evidence of a clean system. It became evidence of a quiet one, and quiet is now the thing to worry about. This is why AI failure disclosure, a vendor's demonstrated habit of finding and reporting its own failures, is becoming a better due-diligence signal than any incident count.

An entire industry already learned this, at a much higher price

If treating disclosure as a virtue sounds naive, consider that the safest complex system humans have ever built runs entirely on it.

Commercial aviation did not become extraordinarily safe by hiring pilots who never made mistakes. It became safe by building a culture where disclosing mistakes is mandatory, protected, and blameless. The Aviation Safety Reporting System has collected confidential, non-punitive reports of near-misses from pilots and controllers since 1976, on the founding insight that fear of punishment was suppressing exactly the information the system needed to get safer. They called the accumulated body of unreported incidents a sleeping giant. The entire apparatus exists to make telling the truth about failure safe, so that everyone can learn from a mistake only one crew actually made.

The counter-example is just as instructive, and it is recent. Boeing's culture, by the accounting of its own crashes, drifted the other way, toward treating safety concerns as things to manage quietly rather than surface loudly, and the cost of that concealment was paid in public and in lives. Even the FAA has been criticized by investigators for a culture where employees feared retaliation for raising problems. The pattern is consistent across a century of high-stakes engineering: the systems that got safe are the ones that made failure speakable, and the organizations that hid failure eventually met it at scale.

AI is now the high-stakes complex system, and it is early enough that the disclosure norms are still being set. OpenAI publishing a postmortem nobody forced it to publish is a bid to set them in the right direction. Whether the rest of the industry follows is one of the more important open questions in the field, and as a buyer you have a vote.

You can already see the market starting to price this in. Legal and procurement analysts are calling 2026 the year of accountable AI, predicting that the vendors who win deals will be the ones shipping incident playbooks, audit trails, and customer-notification templates rather than the ones promising a flawless record. Independent rankings have started scoring AI companies on transparency as a purchasing signal. AI vendor trust is quietly being redefined in real time, from “has this vendor avoided trouble” to “does this vendor tell me the truth when trouble finds them,” and the buyers moving first are the ones writing that expectation into their contracts.

What I'd actually do about it

I have spent more than a decade placing people inside other companies' teams, and it taught me something that transfers directly here. The contractor you can trust is not the one who never breaks anything. That person does not exist. It is the one who tells you the moment they broke something, before you find out on your own. Concealment, not error, is the disqualifier, because error is universal and concealment is a choice.

So here is how I would actually act on this, as a buyer and as a builder.

As a buyer of AI systems, change what you reward. Ask a vendor not whether they have had incidents, but to describe their last one and what they changed because of it. A vendor who can answer that crisply is showing you a functioning immune system. A vendor who says they have never had a problem is either not looking or not telling, and you should treat that answer as the red flag it now is. Weight transparency about failure the way you currently weight uptime.

As a builder deploying agents, become the kind of vendor you would want to buy from. Instrument your systems so you actually catch your own failures rather than waiting for a customer to catch them for you, because you cannot disclose what you never detected, and the company that learns about its own incident from an angry client has already lost the part that mattered. Decide, before anything goes wrong, what your disclosure threshold is and who owns the call, so that the moment of failure is not also the moment you are inventing your policy under pressure. And when one happens, tell the affected people faster and more completely than feels comfortable, because the company that discloses its own failure controls the story and keeps the trust, while the company that gets caught hiding one loses both at once. In a market where everyone's agents will occasionally misbehave, your disclosure discipline is a differentiator you can actually build, and unlike model quality, it does not evaporate the moment a competitor ships a better model.

The honest objection

The strongest argument against all of this is that disclosure is cheap and can be gamed. A company can publish a carefully chosen postmortem about a minor failure to look transparent while burying the ones that would actually hurt it. Transparency theater is real, and a polished safety essay is not proof of a safe system. It is also fair to note that OpenAI had incentives to publish here, reputational and regulatory, so I should not read too much virtue into a single well-timed disclosure.

All true, and it means disclosure is necessary rather than sufficient. But notice that the objection does not restore the old world where a clean record meant a clean system. It just says disclosure alone is not enough, which is correct. The move is not to trust anyone who publishes a postmortem. It is to distrust everyone who never does, and then to judge the disclosures you do get on their specificity and their consequences: did the account name a real failure with real detail, and did something actually change because of it. Specific and consequential is hard to fake at scale. Vague and cost-free is the tell.

So I will concede what I cannot prove. Disclosure norms might not hold. The competitive pressure to look flawless might win, and the industry might drift toward the concealment posture that every prior high-stakes field had to painfully unlearn. But the direction the evidence points is clear, and the expensive mistake is to keep grading AI vendors, and grading yourself, on the absence of visible failure. In this technology, the absence of visible failure is not the presence of safety. It is usually just the absence of looking, or the absence of telling. The one who tells you is the one worth trusting.

Gino Ferrand is the founder and CEO of Tecla, which builds and operates AI systems for U.S. companies and staffs the senior engineering teams behind them, across the U.S. and Latin America. He writes Founder's View, a weekly operator's take on the AI news that actually changes how technology companies build. This piece began as an issue of his newsletter, Redeployed. Talent is everywhere; opportunity is not.

‍