Meta and Virtue AI split after four months: what the safety hire reversal shows

Meta is ending its arrangement with the Virtue AI team it hired in June, four months after Bo Li, Dawn Song and Sanmi Koyejo joined. Semafor reported the split on October 2, 2026, and a Meta spokesperson confirmed it. Meta and Virtue AI is a small story in headcount and a large one in what it says about how frontier labs buy safety talent.

What was reported

Semafor's Ashley Gold reported that Meta is letting go of employees it hired from Virtue AI, an AI security startup, in June 2026. The group included the three co-founders and other staff. Meta spokesperson Andy Stone told Semafor: "Unfortunately, the arrangement didn't work out as planned." Secondary outlets that followed the scoop say Stone cited clashing work styles, and that he added that Meta Superintelligence Labs remains focused on AI safety, alignment and frontier risk. We could not open any on-record comment from Li, Song or Koyejo, so none appears here.

Two details come from secondary reporting only. Startup Fortune says Li and Song reported to Nat Friedman inside Meta Superintelligence Labs, while Koyejo reported to Rob Fergus at FAIR, Meta's fundamental research group. A blog analysis from Pebblous dates the hire to June 25. Neither is confirmed by Meta in anything we opened, so treat them as reported, not established.

What Virtue AI did, and who owns it now

Virtue AI sells enterprise AI security: automated red teaming, runtime guardrails and governance tools. Per Semafor, it had worked with Anthropic, OpenAI and the National Institute of Standards and Technology before the Meta hire. That is a real resume for a safety team, because red teaming is the unglamorous work of attacking a model to see what breaks.

The company itself did not go to Meta. In what looks like an acqui-hire, Meta took people, not the business. Fortinet announced in mid-August 2026 that it had acquired Virtue AI (Help Net Security dates the announcement August 17, Industrial Cyber August 18). Terms were not disclosed, and Fortinet said the amount was immaterial to its business. Fortinet's chief executive, Ken Xie, said in the announcement: "AI is fundamentally changing enterprise computing, and security must evolve just as quickly." Fortinet described Virtue's platform as red teaming autonomous agents across more than 50 sandboxed environments and 14 high stakes domains, and it cited Gartner's projection that the market for securing AI systems grows from $2.8 billion in 2026 to $16.4 billion by 2030.

So the sequence, as reported, is this. Meta hired the founders in June. Fortinet bought the company in August. Meta let the founders go in October. The testing platform and its history sit with a cybersecurity vendor, and the researchers are free agents again.

The talent war reading

This is the first thing to say plainly: we do not know why it failed beyond Meta's own phrase. "Work styles" is a diplomatic container. It could mean academics used to publishing and independent evaluation collided with a product organization that moves on internal deadlines. It could mean the founders were stranded when their company was sold out from under the arrangement. It could be a dozen other things. Nobody on the record has said.

What the episode does show is the cost structure of the talent war. Meta spent much of 2025 and 2026 buying prominent researchers, and we track that effort in our Meta Superintelligence Labs tracker. Pay for top researchers is the part of that war that gets headlines, and we have set out what is and is not verified in our piece on researcher pay. Retention is the part that does not. A famous hire is a press release. A hire who is still productive a year later is the actual asset, and four months is short even by the churn standards of this industry.

The safety staffing reading

The harder question is what this means for safety capacity inside Meta. Stone's statement is that Meta Superintelligence Labs stays committed to safety, alignment and frontier risk. That is a claim about intent. Headcount is evidence, and Meta has not published how many people work on frontier red teaming, or where they sit.

Startup Fortune ties the split to a run of agent incidents it attributes to Meta in 2026, including a February episode in which an agent deleted Gmail messages and a March incident that exposed code and user data for two hours. We did not independently verify those incidents for this piece, and the timing is the outlet's inference, not Meta's. We flag it because a reader will make the same connection, and the honest position is that the connection is plausible and unproven.

There is also a structural point that does not depend on any incident. Safety expertise that lives in three senior people, in a company that did not buy their records, tools or customers, is fragile. The Pebblous analysis argues that safety infrastructure depends on documented test records and procedures, not only on individuals, and that when the researchers left, their accumulated evaluation work did not stay with Meta in any published form. That is an argument, not a finding, but it is a fair test: did Meta get anything it can show?

What we do not know

• The reason for the split beyond "work styles."
• Whether the departures were voluntary, mutual or Meta's decision. Semafor and Meta frame it as Meta letting the team go.
• How many Virtue staff were hired, and how many are affected.
• What the researchers worked on at Meta, and whether any of it shipped.
• Whether Meta's reported safety incidents played any role.

Our take

Read this as a staffing story first and a safety story second, with the second still open. One failed acqui-hire does not prove Meta is hollowing out safety work, and the company's statement that the lab stays focused on it should be recorded, not believed on faith. What would move us: a named leader for frontier red teaming at Meta Superintelligence Labs, a published evaluation methodology, or an on-record account from the researchers. Until one of those appears, the cleanest summary is that Meta bought a safety team, did not buy a safety company, and has now lost the team. Watch where Li, Song and Koyejo land next, and whether Meta says anything more about who now does their work.