Meta is letting go of employees it recruited from AI security startup Virtue AI after they joined Meta Superintelligence Labs in June, spokesman Andy Stone confirmed on 2 October. Stone told Semafor the arrangement 'didn't work out as planned', with the publication reporting that Meta attributed the split to clashing work styles.
The split became public three days after Meta joined five other major AI companies in signing the voluntary White House Accord on Super Intelligence. The agreement calls on companies developing frontier models to police their systems through internal controls, independent external evaluation and board-level oversight.
There is no evidence that the Virtue AI employees were removed for raising safety concerns or because of Meta's support for the accord. Meta also says its broader work on AI safety, alignment and frontier risk remains intact.
Even so, the timing puts a sharper focus on a central question behind voluntary AI governance: how durable are internal safety structures when specialist teams can change quickly?
Meta Hired Virtue AI Founders to Strengthen Security
Axios reported on 25 June that Meta Superintelligence Labs was hiring Virtue AI co-founders Bo Li, Dawn Song and Sanmi Koyejo, along with other members of the startup.
Virtue AI developed enterprise security tools covering automated red-teaming, runtime guardrails and AI governance. The company had also worked with Anthropic, OpenAI and the US Commerce Department's National Institute of Standards and Technology.
An internal Meta message reported by Axios said that as the company shipped AI products to billions of people and built more capable agents, keeping those systems safe, reliable and trustworthy was 'foundational'.
On 2 October, Semafor reported that Meta was letting go of employees hired from Virtue AI, with Stone citing clashing work styles as the reason for the split.
'Unfortunately, the arrangement didn't work out as planned,' he said.
Stone added that Meta Superintelligence Labs remained focused on AI safety, alignment and frontier risk.
Zuckerberg Says AI Labs Have Incentives to Act Safely
The departures draw added attention because Mark Zuckerberg has argued that AI companies already have strong incentives to manage safety themselves.
In a 15 September post on X, Zuckerberg said competition and liability give AI laboratories reasons to act individually on safety. He said every lab has the responsibility and incentive to move at the pace required to train its models safely and can take its own steps to ensure that happens.
That approach was reinforced on 29 September when Meta, OpenAI, Anthropic, Google, Nvidia and xAI signed the White House Accord on Super Intelligence, subtitled the Joint Commitment on Frontier Responsibilities.
The accord sets out four layers of safeguards: robust internal controls, an empowered internal oversight team, an independent external auditor or evaluator, and an independent board committee. The participating companies also agreed to meet regularly to establish safety standards and best practices.
The agreement is voluntary rather than a legally enforceable regulatory regime. Trump described it as 'morally binding'.
The Virtue AI departures became public three days after the accord was signed. That sequence does not establish a causal link, but it provides a timely example of the practical questions surrounding industry-led safety commitments.
Meta Says Its Broader Safety Framework Remains Intact
Meta has also formalised safety processes for its most advanced models.
On 8 April, the company published an updated Advanced AI Scaling Framework covering severe and emerging risks including cybersecurity, chemical and biological threats, and loss of control.
Meta also introduced Safety & Preparedness Reports designed to detail risk assessments, evaluation results, deployment decisions and known limitations.
Those measures matter because the departure of one group does not mean Meta's wider AI safety programme has disappeared. The company's stated explanation remains that the Virtue AI arrangement failed because of incompatible working styles.
Public confidence in industry self-policing, however, remains limited.
A Reuters/Ipsos poll conducted from 17 to 20 September found that 73% of US adults worried AI companies had not gone far enough to prevent AI from causing serious harm to society. Some 55% said slowing AI development would be a good thing.
The Bigger Test for AI Self-Regulation
For Meta, the issue is larger than why one group of employees left.
Self-regulation depends on companies maintaining credible controls, oversight and specialist expertise even as they compete to build more capable systems. Meta says those structures remain in place, and the Virtue AI departures do not prove otherwise.
But Meta is also publicly backing a system in which companies take primary responsibility for developing their own advanced AI systems safely. Changes involving specialists recruited specifically to strengthen AI security will therefore attract scrutiny.
The Virtue AI arrangement lasted about four months, according to Semafor's account of the hires and departures. Whether Meta's wider model of AI self-regulation proves more durable will take much longer to answer.
