Dawn Song (@dawnsongtweets), Bo Li (@uiuc_aisecure) and Sanmi Koyejo (@sanmikoyejo) are joining Meta (@Meta) Superintelligence Labs, a same-day talent move that puts three academic AI security founders inside Mark Zuckerberg's most important AI organization.
Dawn Song on X The move was flagged in an X post and detailed Thursday by Axios, which reported that Meta is hiring the three Virtue AI founders plus other members of the San Francisco AI security startup's team. Axios said Li and Song will report to Meta Superintelligence Labs executive Nat Friedman, while Koyejo will report to Rob Fergus, the head of FAIR, Meta's fundamental AI research lab. The terms were not disclosed.
That matters because this is not being presented as a clean acquisition. Axios described a hiring arrangement involving three founders and other Virtue AI team members, not a purchase of the company, its IP, its contracts or the whole workforce. Virtue AI's own team page also lists Carlos Guestrin (@guestrin) as founder and chief scientist; Axios did not name him among the Virtue AI founders joining Meta.
The founder bet Meta is buying
Song is the name that makes the hire read differently from a routine AI acqui-hire. She is a UC Berkeley computer science professor and co-director of the Berkeley Center for Responsible Decentralized Intelligence, with research spanning AI safety and security, agentic AI, deep learning, security, privacy and decentralization. Berkeley lists her as a MacArthur Fellow, ACM Fellow and IEEE Fellow, and says she earned her Ph.D. in computer science from UC Berkeley and her master's in computer science from Carnegie Mellon.
Li, Virtue AI's founder and CEO, brings the operator seat. The University of Chicago lists her as a research associate professor of computer science and data science whose work focuses on trustworthy machine learning, including robustness, privacy and generalization. She earned her Ph.D. in computer science from Vanderbilt, then worked as a UC Berkeley postdoctoral researcher with Song before joining UIUC in 2018.
Koyejo brings a third academic pole. Stanford lists him as an assistant professor of computer science and leader of the Stanford Trustworthy AI Research lab, which works on AI evaluation science, algorithmic accountability and privacy-preserving machine learning. On Virtue AI's site, he is listed as founder and chief AI officer.
Virtue AI was built around a straightforward thesis: enterprises will not deploy autonomous AI systems at scale unless they can test them, govern them and stop unsafe actions in production. The company's platform spans red-teaming, runtime guardrails, compliance workflows and agent security. Its product pages describe agent action guardrails, agent red-teaming, shadow-AI detection, real-time multimodal guardrails and compliance mapping across frameworks including GDPR, FINRA, HIPAA and the EU AI Act.
Why Meta wants this team now
Meta's AI problem is no longer only model quality. It is distribution. The internal memo quoted by Axios put the point plainly: as Meta ships AI products to billions of people and builds more capable agents, safety, reliability and trustworthiness become foundational. That is the argument for bringing in founders who have spent their careers studying how models fail under adversarial pressure.
The hire also fits Meta's broader reorganization around superintelligence. RuntimeWire reported this month that Zuckerberg's AI reset has moved beyond model ambition into the harder question of commercialization in Zuckerberg's $14 billion AI reset now needs customers. We also noted the internal strain of Meta's AI reshuffling in Boz tells Meta employees its AI reorg rollout was 'atrocious'. Hiring a tight security research team is the cleaner part of that strategy: buy scarce judgment, not another sprawling org chart.
Agent security is the sharp edge. Chatbots can say the wrong thing. Agents can call tools, touch data, trigger workflows and make chained decisions across enterprise systems. That changes the risk surface from content moderation to prompt injection, data leakage, unauthorized tool use, compliance violations and autonomous behavior that operators cannot fully inspect after the fact.
Virtue AI has been selling directly into that fear. In April 2025, Axios reported that Virtue AI had raised $30 million across seed and Series A funding, led by Walden Catalyst Ventures and Lightspeed Venture Partners. The round was not new, and its valuation was not disclosed. Axios reported at the time that Virtue AI counted Uber and Glean among customers, along with companies in finance, healthcare and IT and several frontier labs.
The unanswered question is what remains of Virtue AI after the move. The company had framed itself as an enterprise control layer for AI deployment, not a research lab waiting to be absorbed. Its site still describes products for financial services, healthcare, insurance, government and Fortune 500 customers. But when a young security startup loses three named founders and other team members to a platform company, the market reads the same signal either way: frontier labs and large vendors want AI safety talent before the agent stack hardens around someone else's controls.
For Meta, the hire gives Superintelligence Labs people who know the failure modes from both sides of the table: academic evaluation and enterprise deployment. For Virtue AI's founders, it moves their work from selling controls to companies deploying AI into one of the companies setting the pace for what agentic AI will become.