Hired in June, gone in October
Meta has parted ways with the employees it brought in from the AI security startup Virtue AI in June, Semafor reported on Friday. The June group included Virtue AI co-founders Bo Li, Dawn Song and Sanmi Koyejo, along with other team members.
Meta spokesperson Andy Stone told Semafor that “unfortunately, the arrangement didn’t work out as planned”, citing clashing work styles, and said Meta Superintelligence Labs remains focused on AI safety, alignment and frontier risk.
Four months is a short tenure for an acqui-hire. It is not long enough for a research agenda to produce much, and it is long enough for the acquiring company to have formed a view.
What the team did before
Virtue AI specialised in AI security and red teaming — the practice of attacking a model deliberately to find what it will do under adversarial pressure. The company had previously worked with Anthropic, OpenAI and the Commerce Department’s National Institute of Standards and Technology.
That is an unusually well-connected client list for a startup of its size, and it is the capability Meta was buying. Red teaming is one of the few safety practices that produces findings a product organisation can act on directly, which is also what makes it uncomfortable to host.
Neither Meta nor the researchers have said publicly what the clashing work styles were, and Semafor’s account rests on Meta’s own characterisation. The researchers have not commented.
The context around it
The departure lands in a week when safety staffing has been visibly unstable at more than one lab. OpenAI dismissed three safety researchers over allegedly sharing confidential information with an outside organisation, and the person who ran its safety transparency work resigned days later.
Acqui-hires of this shape have become the standard way a large lab buys a research capability it cannot recruit piecemeal, and they fail quietly more often than the announcements suggest. What makes this one legible is the subject matter: a team hired explicitly to find problems, let go by the company whose problems it was hired to find.
Those are different companies and different circumstances, and it would be wrong to read a single trend into them. What they do have in common is that in each case the function being reorganised is the one whose output is a list of things the company’s products get wrong.
Meta’s own position is that the work continues inside Meta Superintelligence Labs. That is a claim that can be checked over time rather than this week: the measure is whether Meta publishes adversarial evaluations of its own models, and who signs them.
