The view from the top of the policy machine is a landscape of blinking screens and silent corridors. From the seventh floor of the Eisenhower Executive Office Building, the White House complex looks like a grid of controlled motion. But the people who understand the code beneath that grid are walking out the door.
Thomas Lind, the head of policy at the Office of the National Cyber Director, is leaving. [1] He is the third senior AI adviser to exit the Biden administration in as many weeks His departure, like the ones before it, is framed as a personal choice — time with family, new ventures, the natural rhythm of government service. But when the people who know how the machine works start leaving before the machine is built, you have to ask: what are they running from, or what is being left behind?
The answer is a specific kind of expertise that the White House cannot afford to lose. Lind came from the National Security Agency. He understood the inside of a zero-day exploit, the kind of vulnerability that lets an attacker walk through a locked door without touching the handle. He was one of the few people in the building who could look at a frontier AI model and see not just a tool, but a weapon that could find every unlocked door in America’s digital infrastructure and open them all at once.
The executive order on AI security, released in October 2023 after months of delay and internal fighting, was supposed to be the guardrail It tasks the ONCD with leading the federal response to models that can autonomously discover software flaws. But the people who wrote that order are now gone. Lind’s deputy, Alexandra Seymour, left last week. Sriram Krishnan, the senior policy adviser for AI, announced he would depart by the end of June. The office that is supposed to implement the policy is bleeding the people who understand the policy.
This is not a story about politics. It is a story about what happens when an organization tasked with understanding a fast-moving technology loses the people who can actually think about it. The ONCD is authorized to staff up to 75 people. It currently has about three dozen. Its director, Sean Cairncross, came from a background in law and Republican Party operations, not cybersecurity. His chief of staff and deputy chief of staff also lack technical backgrounds. The people who could translate the language of machine learning into the language of regulation are leaving, and the people who remain are asking questions they cannot answer.
The research team that first raised the alarm about these models was small. [2] They worked at Anthropic, a company founded by former OpenAI researchers who believed that safety had to be built in from the start, not added as an afterthought. In early 2024, they announced that their latest model could find security vulnerabilities in code with a success rate that shocked even them. It was not just that the model was good at finding bugs. It was that it found bugs no human had ever seen, in code that had been reviewed by dozens of experts. The model saw the cracks in the wall that everyone else had walked past.
The moment of discovery came late on a Tuesday night. The lead researcher, a woman in her early thirties who had spent years studying how neural networks generalize from examples, ran the model against a set of open-source libraries used by millions of people. She expected it to find a few known vulnerabilities. Instead, it found eleven zero-days — flaws that had never been reported, never been patched, never even been noticed. She sat in the dark lab, the glow of the screen reflecting off her glasses, and realized that the world had just changed. A machine could now find the cracks in the digital foundation of modern life, and it could do it faster than any human.
The question that followed was simple and terrifying: what happens when this capability is not just in the hands of a safety-conscious company, but in the hands of anyone who can run the model? The answer is that every system becomes vulnerable. Every hospital network, every power grid, every financial exchange — they all have cracks. They always have. The only thing that protected them was the time it took for humans to find those cracks. Now that time has been compressed to near zero.
The White House response was slow. The executive order was delayed by infighting between those who wanted to regulate the technology heavily and those who wanted to let it develop freely. The AI industry lobbied hard against the earlier draft, and the final version was weaker than originally planned. By the time it was signed, the people who understood what was at stake had already started to leave.
Lind’s departure is the most significant because of what he represented. He was not a politician or a lawyer. He was a technician who had learned to speak the language of policy. He could sit in a meeting with Google engineers and understand what they were saying, and then turn around and explain it to a senator. [3] He could look at a model’s training data and know whether it was safe to release. He could look at a vulnerability report and know whether it was a real threat or a false alarm. That kind of person is rare in any organization. In a government that is already short on technical talent, it is irreplaceable. The loss of such expertise leaves a gap that cannot be filled by policy alone.
The ONCD is now searching for replacements. A White House official said the office is “actively” looking to hire qualified candidates. But the people who understand AI security well enough to advise the president are the same people who are being offered millions of dollars by private companies. They are the same people who are founding their own startups. They are the same people who are leaving the White House to start outside institutions that will still influence policy, but from a distance. The government cannot compete on salary, and it cannot compete on speed. It can only offer the chance to serve, and that is not enough when the stakes are this high.
The historical pattern is clear. Every time a transformative technology emerges, the government struggles to keep the people who understand it. During the Manhattan Project, the scientists stayed because they believed the alternative was unthinkable. During the early days of the internet, the engineers stayed because they believed in the dream of a connected world. But AI is different. The people who build it are not sure what it will become, and they are not willing to wait for the government to figure it out. They are leaving to build the future on their own terms, and the government is left holding the policy papers.
The research team at Anthropic continues to work. They have not released the model publicly, and they are trying to understand how to contain it. But they know that other companies, other countries, other actors are working on the same problem. The genie is not going back in the bottle. The only question is who will control it.
The door closes behind Thomas Lind as he walks out of the White House for the last time. He does not look back. He has a family to spend time with, and he has done what he could. But the building behind him is quieter now, and the people inside are looking at the executive order on their desks, trying to figure out what it means. They are reading the words, but the person who wrote them is gone.
The hand that reaches for the next briefing is not the same hand that wrote the last one. It is a hand that knows the law, but not the code. It is a hand that can sign a document, but not read a model’s output. It is a hand that is reaching into the dark, hoping to find a switch, but not knowing where the light is.
Sources
1. Office of the National Cyber Director
2. Anthropic
3. Google
4. White House