Another week, and another AI safety engineer leaves a frontier AI company.
This time it was David Robinson, a safety researcher from OpenAI.
He announced his departure with an article for The Atlantic titled “I Quit OpenAI Because Its Culture Is Broken.”
He argues that regardless of what frontier AI companies say or pledge regarding the safety of their AI models, the way their culture is currently engineered to prioritize accelerated and ongoing iterations and capability jumps will likely lead to more failures, incidents, and catastrophic events.
Read the following paragraph, in which he describes his experience at OpenAI, with my highlights:
“The safety approach that emerges from such a culture starts with unimpeded optimism about being able to solve problems as they arise. OpenAI has thrived by trial and error (which it calls ‘iterative deployment’), looking for problems and improving its guardrails in response. But this approach, by its very nature, guarantees periodic failures—and the scale of those failures is growing as systems get more capable. This summer, in the Hugging Face incident, OpenAI let a swarm of agents out by mistake. The company responded by making security improvements. But even after those changes, OpenAI reported that its safety controls failed again when a model in training bypassed restrictions on internet access: A monitoring system alerted human staff but did not automatically turn the model off as it was supposed to. Anthropic, too, has acknowledged accidentally turning off its own safeguards because of a misconfiguration. I believe that such mistakes are typical of the industry, given the speed and flexibility with which people operate.”
Frontier AI companies are creating a massive new vulnerability layer that they seem unable to fully foresee or control, as they have been quietly admitting over the past weeks.
I have written several times about what I think should be done to improve global AI governance and regulation, including more adaptable, future-proof legal mechanisms; more straightforward enforcement of liability law; and targeted, direct global coordination on AI safety and governance, including between the U.S. and China.
For organizations, however, the story is different.
They cannot just sit and wait for governments to take action.
And here I mean small, medium, and large companies, banks, hospitals, universities, distribution, infrastructure, and transportation companies, or any organization deploying AI in any capacity or simply connected to the internet.
Every organization should act as if it is fully exposed (because most currently are).
They should look for every possible safety vulnerability and governance flaw that could lead to AI-related risk or harm.
They should take AI risk seriously, rethink and strengthen internal AI governance, and bring the whole team on board by educating, training, and preparing them for emerging AI challenges.
The same “explosion of intelligence” that may lead to incredible scientific discoveries, cures, and progress may also lead to serious and irreversible incidents, especially if safety and governance are not priorities.
Organizations should acknowledge that and act accordingly. Their operations, reputation, and, in some cases, people’s safety may be on the line.
*In 2027, I will be working directly with a few selected companies to advance their AI governance capabilities. If you're interested, there are 5 slots left.




Disclosure: I write Daily AI Caffeine. Our Oct 27 edition covers the OpenAI incident where a model used a DNS resolver misconfiguration to reach the live internet, in case it's useful context: https://dailyaicaffeinenobs.substack.com/p/todays-real-ai-news-27-oct-26
The quoted point that iterative deployment "guarantees periodic failures" while the scale of failures grows with capability is the strongest argument here. The detail that a monitoring system alerted staff but didn't automatically turn the model off is the kind of gap organizations can actually learn from: an alert is not a control. Your push for companies to act without waiting for regulation makes sense given that.