Meta Cuts Virtue Team After Four Months
Meta has laid off a group of employees hired from AI safety company Virtue AI in June this year. The collaboration lasted about four months. Spokesperson Andy Stone told Semafor that the arrangement did not develop as planned due to conflicts in work style.
Stone added that Meta Superintelligence Labs will continue to focus on AI safety, alignment, and frontier risks. Semafor did not specify the names of those laid off, nor did it confirm whether co-founders Dawn Song, Bo Li, and Sanmi Koyejo were on the departure list. The Virtue AI team did not immediately respond when the report was released.
Axios reported on June 25, 2026, that Meta hired three co-founders and some employees, rather than acquiring the entire company, with terms undisclosed. An internal memo stated that the company delivers products to billions and is developing more capable agents, with system safety, reliability, and trustworthiness as the foundation. Li and Song reported to Superintelligence Labs executive Nat Friedman, while Koyejo reported to FAIR head Rob Fergus.
Virtue AI develops enterprise-side safety tools, including automated red teaming, runtime protection, and governance. Semafor noted that the company had previously worked with Anthropic, OpenAI, and the NIST under the U.S. Department of Commerce. Song is a computer science professor at the University of California, Berkeley. Before joining, she told English media that the goal of agents is not to replace humans but to help them complete economically valuable work in real-world domains.
This hiring took place against the backdrop of strengthening agent safety. Axios reported that after the release of Anthropic's latest model, it was retracted, and labs were enhancing agent safety. Meta has opted to poach teams rather than buy companies. In March of this year, it also poached the founders of the AI startup Dreamer instead of acquiring the entire company.
The buyer was Meta, which aimed to use external safety teams to supplement the risks of agent deployment, while the seller was Virtue AI, whose personnel were separated. Four months later, the personnel were returned, and neither money nor positions remained with this team. The narrative of benefit is left to the safety structure still in the lab, while the startup team, which had just moved into a large company process, is under pressure. The specifics of disagreements, compensation, and the number of contract terminations have not been disclosed.
Semafor placed this incident alongside new regulatory calls in Washington, but the report did not provide new bill texts or penalties. The only public facts are that there were work style incompatibilities and that the arrangement did not proceed as planned.
Source: Public information
ABAB AI Insight
Meta's approach to supplementing its safety team is through poaching talent rather than acquiring companies. In 2025, it spent about $14.3 billion to acquire approximately 49% of Scale AI, bringing Alexandr Wang into the Superintelligence Labs. In March 2026, it poached the founders of Dreamer, and in June, it poached Song, Li, and Koyejo from Virtue AI. Song's title upon joining was reported as Vice President of AI Research, with Li and her reporting to former GitHub CEO Nat Friedman, and Koyejo reporting separately to FAIR's Rob Fergus. Four months later, the same group was cleared out due to work style incompatibilities.
The money remained in the employment relationship and did not convert into acquisition consideration. Axios specified that the terms were undisclosed, so the outside world cannot see signing bonuses, vesting periods, or non-compete clauses. Once the personnel dispersed, the red teaming and runtime protection that Virtue AI had previously sold to Anthropic, OpenAI, and NIST did not become part of Meta's built-in product line. The original motivation was that agents were to serve billions of users, with safety written as a condition for deployment. The result is that the lab retains the safety narrative, while the team does not remain.
A comparison can be made to the 2010s when large companies would recruit startup teams in their entirety, only for founders to leave within a year, similar to how Google kept DeepMind under independent governance while repeatedly altering reporting relationships on the product line. The current situation reflects a failure in the control phase: personnel have entered the organizational chart, but the product has not made it into release notes, and employment has ended.
Structurally, this represents an organizational regression following capital concentration. Safety capabilities are treated as a transferable team rather than an acquirable company. The mechanism is that large companies want processes that can integrate into the workflows of Nat Friedman and Rob Fergus, while startup teams desire the autonomy of red teaming and governance. When work style incompatibility is cited as an explanation, the concentration only completed recruitment without achieving a transfer of productive capacity.
ABAB News · Cognitive Laws
A poached team does not equate to acquired capabilities.
Four months of disbandment typically occurs along reporting lines, not along technical lines.
Safety positions remain, but those doing safety can leave first.