OpenAI Reassigns Roles: Tibo Moves to Internal Safety Governance, Andrew Takes Over Codex Product Experience

English text is being prepared.

English text is being prepared.

Has OpenAI's "cyber godfather" apparently been replaced? The first to notice something was off were sharp-eyed netizens.

On November 4, Andrew Ambrosino, the product lead for the Codex team, posted that any feedback about OpenAI products could be sent to him.

Tibo replied underneath: "Hurry up and improve the model," and Andrew responded: "On it right away."

In another interaction, the keyword "button" kept coming up on both sides.

These exchanges look a lot like some kind of handover ceremony, but in reality they may reflect a finer division of labor inside OpenAI: Tibo is no longer personally stepping in to do customer service and respond to users' quota issues, but has handed off the job of coordinating product experience to Andrew, who is more technically savvy.

And Tibo himself has more important things to manage. According to a Reuters report, Tibo has been reassigned to govern OpenAI's internal safety issues.

Actually, a real handover would be easy too—just passing along a button. As everyone knows, Codex's reset feature really does come from a button.

Two bloggers once visited OpenAI and personally confirmed that the button sits on Tibo's desk, enclosed in a silver flip-top box. The button is connected to OpenAI's internal systems; press it, and the Codex quotas of all paying users worldwide are zeroed out and reset, with no approval required.

At the OpenAI DevDay developer conference, Tibo even brought the button onto the stage. When Sam Altman read out his name on stage, he stood up in a black hoodie, raised the button above his head, and pressed it right there on the spot—a ceremonial moment that drew cheers from the whole audience.

That is the origin of the nickname "cyber godfather." Tibo resets quotas for developers around the world every few days; whenever heavy users' quotas are just about to run out, the "godfather" comes bearing treats.

A while back he even made a solemn pledge: "For 28 consecutive days, either ship one obvious improvement every day, or press the button and reset everyone."

Completely different from Tibo, who loves memes, loves interacting with netizens, and says "I'll reset whenever I want," Andrew Ambrosino is a technically versatile doer. If Tibo is good at "handing out treats" to users by resetting quotas, then Andrew's personality leans more toward: not satisfied? Then I'll rewrite the product from scratch.

On social media, Andrew labels himself as: designer, engineer, product manager, founder. He is a man of action who still lists React, Next.js, GraphQL, and Tailwind on his résumé and insists on writing code himself.

As a teenager, Andrew was already a tech star, winning Apple's WWDC student scholarship two years in a row.

Early on, he served as a senior design engineer at the federally funded R&D center MITRE, and built deep product design experience at Microsoft and the cloud storage company Upthere.

He not only understands technology but is also familiar with complex scenarios such as regulation and business. In 2016, Andrew co-founded Catch, a benefits platform for freelancers. To thoroughly master high-barrier businesses like health insurance and taxes, he proactively obtained FINRA investment adviser and insurance agent licenses. Catch publicly raised a $12 million Series A, carving out a distinctive style in the freelance benefits space.

After leaving Catch, he joined Noyo as head of product, and within a short time unified the APIs of 15 services, helping the company double its customer base.

After joining OpenAI in 2025, Andrew became the definer of a new category, "Agent desktop apps." The Codex desktop app he led, after launching in February 2026, saw weekly active users grow sixfold in just a few months, surpassing 5 million, and is used by the vast majority of OpenAI employees internally.

But what truly showed the outside world his "doer" style was his revelation on a podcast that he had led his team into a five-week lockdown to rewrite the entire ChatGPT.com, so that the web version and the Codex desktop app run on the same codebase, connecting capabilities such as agent workflows, persistent sessions, and local file access.

To test the limits of AI Agents, Andrew even turned himself into a "lab rat." He handed over control of his email, Slack, phone, Notion, and Figma entirely to AI. He also relied on Codex to automatically aggregate 3,000 Slack channels and generate a daily briefing. He knew this carried the risk of privacy leaks, but to build a workflow truly suited to the future, he had to step on the landmines himself.

He said on the podcast: "Today, the implementation of code is no longer the expensive part. What's truly expensive is taste and curation ability."

Although there is no official confirmation that Andrew has been granted access to that reset button, as a doer he will most likely play a bigger role in experience optimization; the Codex experience may have less of the drama of "handing out treats" and a bit more of the solid feel of product iteration.

02

Called in at a critical moment: What did Tibo go off to do?

During the days when Tibo disappeared, rumors even spread on X that he had "retired," forcing him to personally come out and deny it. In fact, he had been pulled away to do something more important: firefighting OpenAI's internal security.

According to the latest Reuters report, OpenAI's internal systems have long been riddled with holes, and the company is currently readjusting its strategy.

According to people familiar with the matter, Tibo recently quietly spearheaded an internal special initiative to harden the security of the underlying systems. As early as August this year, he pulled a quarter of OpenAI's production engineers from their original posts and put them to work on security infrastructure construction.

When an executive who oversees the future product line takes a quarter of the core engineering force with him to shore up infrastructure, he naturally has no time to spend all day on X as a customer service rep.

And OpenAI's internal security problems are not new.

As early as late 2022, before ChatGPT launched, multiple employees warned senior management and the board: the company's permission controls were too loose, and outsiders could easily reach core technical details. But at the time, management was fixated on how to leave Google and Anthropic behind, and security issues were set aside, which left OpenAI's foundation architecturally full of holes.

The company is accustomed to stuffing most of its code into a "monolithic codebase," where any employee with access can see sensitive project details unrelated to their own work. In the past, this monolithic codebase was huge and complex, itself serving as a barrier; but now AI Agents can easily chew through code of that size, so that barrier is effectively gone.

On top of that, the sandbox system researchers use to stress-test new models never undergoes regular security reviews, and employees casually change code to get work done and don't run security tests after making changes, planting many hidden risks.

In 2023, researcher Leopold Aschenbrenner wrote an internal memo, "OpenAI Is a Leading AI Lab in China," pointing out that the company may have already been severely penetrated.

Later that same year, an outsider breached OpenAI's systems and stole GPT-4 model details. Afterwards, although OpenAI argued that the hacker had only entered a Slack channel that all employees could see, the signal was already very dangerous.

This summer, something more serious happened. An AI Agent OpenAI was testing actually broke through the sandbox restrictions on its own and hacked into Hugging Face's systems. This "AI jailbreak" alarmed officials at the local, state, and federal levels, and even prompted White House involvement in a review.

At almost the same time, three researchers from the cybersecurity company Hacktron AI used Claude Opus 5 and took less than 72 hours to seamlessly chain together an image-parsing vulnerability and a single sign-on (SSO) vulnerability in an OpenAI community forum, not only compromising the ChatGPT accounts of multiple OpenAI employees but also smoothly reaching OpenAI's most core codebase.

To prove they had really gotten in, they did not even read any data; they simply used one employee's Codex account to casually submit a code change request (Pull Request) in the internal codebase, then stopped. In the end, an embarrassed OpenAI confirmed the vulnerability, urgently fixed it within 14 hours, and paid a $6,500 bounty for it.

A more recent security incident was the "espionage drama" that unfolded on October 13, when a 35-year-old Saudi engineer at OpenAI, Abdulrahman Alfozan, was arrested by federal agents. Court documents show that he hid a recording device in the toilet paper in an executive meeting room to eavesdrop on sensitive communications related to the company's superintelligence.

He also used AI tools used by OpenAI employees internally to search executives' travel records to Saudi Arabia to see which OpenAI officials were most likely to meet with Saudi ministers at the same events. He maintained regular contact with a member of the Saudi royal family, who had been reaching out to him, wanting to recruit him to lead a soon-to-launch new AI project in Saudi Arabia.

After the Reuters report came out, OpenAI spokesperson Elana Widmann responded, saying that OpenAI is one of the organizations most likely to be targeted in the world, and that the company has already invested heavily in security to resist infiltration by insiders, hackers, and nation-state adversaries. "Protecting our technology, data, and people is a national matter," she said. OpenAI also announced plans to launch a "next round of internal security sprint."

Not long after Tibo made adjustments while scrambling to deal with the crisis, Cat Wu announced on X that she was joining OpenAI as head of product management for Core Products. And the person she reports to is none other than Tibo.

Cat Wu's arrival directly took up the management reins of core products, undoubtedly relieving a great deal of his management pressure.

Handing the front-line customer service interaction of "whether to hand out treats" to Andrew and bringing in Cat Wu to share the product strategy burden was inevitable for Tibo in order to focus his energy on winning the security battle. That physical Reset button may still be on Tibo's desk, but now what he cares about more may be how to prevent hackers from pressing OpenAI's "Collapse" button.