Meta and Virtue AI disbanded, and Berkeley Professor Song Xiaodong's team joined only four months later. Due to their different working styles, the AI security cooperation changed.
Original title: Meta parts ways with Virtue AI
Original author: Ashley Gold
Original compilation: Machine Heart
Recruited in June, and by October there are reports of a breakup?
The Virtue AI team of UC Berkeley computer science professor Dawn Song is reportedly seeing a change in its collaboration with Meta, with the reason being "clashing work styles."
Dawn Song enjoys a high reputation in both academia and industry, and is one of the most influential scholars in the fields of global computer security and AI safety. Her 2005 paper proposing "Dynamic Taint Analysis" is a classic paper in the field of computer security.
In the rankings compiled by AMiner, she is the most highly cited scholar in the world in the field of computer security. Her laboratory at Berkeley is known as the Whampoa Military Academy of the computer security field.
In addition to being a professor, Dawn Song is also the founder of Oasis Labs and Virtue AI. According to the introduction on the Virtue AI official website, what they do includes simulating attacks, finding AI weaknesses, and providing protection during system operation while implementing safety and governance rules. In plain terms, they both help AI undergo a "security checkup" and install "guardrails" for it during operation.
In June of this year, Dawn Song announced on social media that she was joining Meta's Superintelligence Labs. According to media reports at the time, the other two founders of Virtue AI, Bo Li and Sanmi Koyejo, as well as more team members, also joined. In other words, this recruitment was Meta bringing in an independent team rather than acquiring the entire company.
For the Virtue AI talent team, joining Meta was because they were already doing AI security testing, red teaming, and post-launch protection for clients, but as a startup they had limited scenarios and resources. Once inside Meta, they could directly extend these capabilities to products used by billions of users, turning from a "security tool" into the security infrastructure of a large platform.
At the same time, Meta also needed them, because it is aggressively pushing AI Agents and superintelligence, and the further it goes, the more afraid it is of models going rogue, data leaks, and being led astray by malicious instructions. So it had to quickly find a top team that both understands research and has actually done security implementation to provide backing for itself in AI safety, reliability, and governance.
All of this seems very reasonable, but in a recent article, foreign media outlet Semafor wrote, "Meta spokesperson Andy Stone told Semafor that the company is laying off a group of employees who were hired from AI safety startup Virtue AI and who had only been on the job for four months since joining in June."
Why did it reach this point so quickly? Meta's explanation is "clashing work styles," and the collaboration did not develop as expected. At the same time, Meta also stated that Superintelligence Labs will still focus on AI safety, alignment, and frontier risks. Of course, this is only the company's side of the story, and the specific disagreements have not been disclosed.
In addition, the report also did not confirm whether Dawn Song and the others had all left, and the team had not yet responded when the report was published.
However, the questions truly left by this change in collaboration are still worth discussing, and the most central one is: after bringing excellent researchers in, how do you get them to play a role within the organization?
A 2019 study published by MIT Sloan found that within the first year after a company is acquired, 33% of acquired employees choose to leave, compared with 12% of regularly hired employees with similar skills and work experience. Although over time these percentages tend to converge, within the three-year window they studied, acquired employees were 15% more likely to leave than new hires.
The article argues that the core reason for the high turnover rate among acquired employees is "organizational mismatch": these people originally joined startups for the entrepreneurial atmosphere, autonomy, flexibility, and risk-taking environment, but after being acquired by a large company, they often have to face more bureaucratic processes and a formal corporate culture, which directly conflicts with their preferences. Moreover, ordinary acquired employees usually cannot choose the buyer, nor can they participate in post-acquisition organizational adjustments. This "having no choice" further intensifies their discomfort, and combined with their inherent preference for building things from scratch, they are even more likely to leave, or even go out and found competing companies.
Returning to this Meta incident, the explanation of "clashing work styles" is clearly unconvincing to many people. As for what deep disagreements exist between the two sides, we still need to wait for an explanation from Virtue AI.