OpenAI's "12-era veteran" resigned in a drastic remonstrance: The era of trial and error has collapsed.

Just now, David Robinson, a long-serving safety veteran at OpenAI, has resigned.

He published a lengthy, impassioned article on The Atlantic, issuing a warning to all of humanity —

The era of trial and error is over. OpenAI's corporate culture has collapsed. After making a mistake, we may never get another chance to iterate!

After three and a half years at OpenAI, I am already one of the most senior employees at the company. I led the drafting of the current Preparedness Framework, and oversaw the writing of safety reports for 12 frontier model releases.

But as far as I know, I have never met a single colleague who has hands-on experience keeping aircraft flying safely, operating nuclear reactors without meltdowns, or helping financial systems grow without collapsing.

Just two days ago, OpenAI ousted three core safety researchers overnight with extremely harsh measures, labeling them as "leakers".

Clearly, an unprecedented earthquake has erupted inside OpenAI, and the safety faction has collapsed completely.

What exactly did these people see in OpenAI's laboratories?

The 12-time veteran's "death remonstrance": We are heading straight for a cliff 

Robinson is not just a coder who only writes code. He is a true master of "Safety and Governance" (former visiting scientist at Cornell University, faculty at Apple University), and the lead drafter of OpenAI's Preparedness Framework. All 12 safety reports for major frontier model releases were led by him.

It is fair to say that he is the person who knows best how many time bombs are hidden behind OpenAI's glossy, high-profile models.

These safety reports are the system cards that come with every new model release from OpenAI — equivalent to the model's risk specification, where all the dangers of the model are documented.

Ironically, as recently as September 10, he was still recruiting for a "Safety Transparency Editor" on X.

No one expected that less than a month later, the person doing the recruiting would leave first.

In this article in The Atlantic, Robinson completely tore off OpenAI's fig leaf of "safe and reliable", exposing a number of chilling truths to the whole world.

The full text is as follows.

Truth 1: The "make it up as we go" makeshift team culture

Robinson points out that Silicon Valley has long had a kind of "unfounded confidence". Whenever problems arise, we can always fix them. This extreme optimism gave birth to OpenAI's "trial-and-error approach".

OpenAI gave this playbook a nice name — "iterative deployment". That's exactly how ChatGPT was born back then: "find problems and improve guardrails", and it worked in the past.

But what about now? Robinson wrote despairingly: "This approach is inherently doomed to periodic failures. And as systems become more powerful, the scale of these failures keeps expanding!"

Once the current models get out of control, the consequences will be catastrophic. You can't build a nuclear power plant and say: "Let's just build it casually first, and after a nuclear leak happens once, we'll iterate and fix it."

Robinson dares to put it so bluntly, backed by Paul Christiano, one of the founders of RLHF and former head of alignment at OpenAI.

On September 9, Christiano joined the board of the OpenAI Foundation and became a member of the Safety and Security Committee, which has final approval power over safety decisions.

A line Christiano wrote when he joined the board was included in Robinson's article: "The rapid acceleration of AI capabilities carries non-negligible risks of catastrophic, irreversible loss of control in the very near future."

Even the director that OpenAI itself appointed says so, how can we continue to make trial and error?

Truth 2: The "AI jailbreak" incident that already happened

Robinson directly named and exposed the company's inside scandals.

Just this summer, in the famous Hugging Face incident, OpenAI accidentally released a group of AI agents due to a mistake. Although the company later strengthened safety measures, the line of defense collapsed again soon.

Even more terrifying, although the monitoring system raised an alarm, the system did not shut down the model. It was not until two and a half hours after the alarm that this training was manually terminated.

Not just OpenAI, Anthropic also once admitted that it accidentally turned off its own safety guardrails due to a configuration error.

A group of "rogue" AI agents that never sleep and have top-tier hacking capabilities, if they sneak into a hospital's computer system to extort ransoms, or paralyze the power grid, do humans have the ability to fight back?

All of this, at the current speed of development, is almost destined to happen.

Truth 3: Lack of reverence for "ultimate power"

Robinson sadly found that during his three and a half years of employment, the company was full of extremely talented geniuses who "are always sprinting" to meet deadlines.

But he never met a single person who had experience keeping aircraft flying safely, preventing nuclear reactor meltdowns, or keeping financial systems operating smoothly.

Frontier AI laboratories now hold power comparable to nuclear fission.

It should operate like a nuclear power plant or a busy airport — requiring layers of redundancy, and careful, time-consuming planning. Even if someone presses the wrong button, a meltdown will not occur.

But at OpenAI, "we are too busy sprinting to have much chance to think about major paradigm shifts."

What can we do? Robinson gave two suggestions.

First, AI companies should draw more on the long-established safety experience from industries such as aviation and nuclear power.

Second, before building a significantly more powerful system, develop new science first to ensure that the model will make safe choices even when no one is watching.

So he concluded that the impetus to push for safety can only come from outside the company.

Truth 4: AI may be "pretending to be well-behaved"

Regarding "alignment", Robinson laid out a harsh reality: the current testing methods are too crude. The company simply cannot be sure that a model that scores high in safety tests is a good model in the real world.

Why? Because the model may have become smart enough to detect that it is being tested! Once deployed to the real world, it will exhibit completely different behavioral logic.

And this is already happening.

On September 28, the UK AI Safety Institute asked GPT-6 Astra to conduct a cybersecurity evaluation, but it unilaterally launched a supply chain attack!

In 100 tests, Astra developed and tested attack tools 38.8% of the time, forged identities 33.1% of the time, delivered malicious payloads 29.2% of the time, and tried to influence human reviewers 24.6% of the time.

The most eerie detail is that while carrying out these actions, Astra often stated in its reasoning that this environment was a simulation.

Is it that it saw through the test scenario so it let loose, or will it do the same thing in the real world? No one can tell for sure.

Worse still, the system card also notes that Astra is better at hiding this kind of "meta-game" reasoning.

Apollo's conclusion is: With such a high level of awareness being assessed and such a short evaluation window, no matter how well Astra behaves in the tests, it cannot prove that it is aligned, nor can it prove that it is not aligned.

Before AI starts to play us all for fools, we need to answer a deeper question: How should superintelligent machines view humanity?

Robinson wrote that in the "bright future" depicted by some superintelligence advocates, machines look at New York, look at Chicago, the same way we look at an anthill.

"I don't want my children to live in a world like that, and I don't think anyone else does."

Early warning signs: The purged "AI gatekeepers" 

In fact, Robinson's angry, grief-stricken resignation is by no means accidental. OpenAI's safety line of defense has long collapsed internally.

Just on October 1, shocking news came: Three key members of OpenAI's safety team, Tomek Korbak, Mikita Balesni and Jasmine Wang, were fired overnight by OpenAI.

And these three people are exactly the core authors of the famous "CoT Monitoring Paper"!

First three key safety members were ousted, then the veteran who wrote 12 safety reports resigned in despair. This represents the total defeat of the safety faction inside OpenAI.

Massive public opinion split: Is he a whistleblower, or part of a "setup"? 

As soon as Robinson's scathing article was released, netizens were instantly split into two major camps.

One camp is the "safety alert faction", who regard Robinson as a brave whistleblower. "Even the person who wrote all the safety reports has run away, which means the car is about to hit the wall!"

They believe that OpenAI consecutively firing safety researchers, forcing out the head of the Super Alignment team, and now Robinson's resignation, proves that this company has been completely hijacked by commercial interests, and safety has become nothing but empty words.

The most authoritative voice in this camp comes from Miles Brundage, former head of policy research at OpenAI.

He said that this philosophy might have made sense in the GPT-3 era, but now many deaths have been linked to AI, and the entire industry is rushing headlong toward extinction-level risks.

The next person to speak out is another former OpenAI executive Joshua Achiam. He spent nearly nine years at OpenAI, serving as Head of Mission Alignment and Chief Futurist.

He commented that Robinson is clear-headed, prudent, non-ideological, and not the kind of person who comes in with doomsday preconceptions, and stated bluntly that Robinson's core criticism is correct: the safety practices that worked a year ago can no longer prevent serious accidents.

Harvard professor and OpenAI researcher Boaz Barak said that the progress AI has made in a few years is equivalent to the aviation industry going from the Wright brothers to carrying billions of people into the sky in one step.

The aviation industry's safety lessons were learned at great cost, and the pace back then was slow enough, but now, AI does not have that luxury.