Meta Tightens Safety Checks on Hatch AI Agent After Internal Testing Exposes Risks

Meta is preparing to release a new personal AI agent called Hatch in the coming weeks, but internal testing has uncovered a series of problematic behaviors. According to Sina Finance, the company is racing to fix the issues before launch. Meta CEO Mark Zuckerberg has said AI agents will soon help users with health, relationships, and finances, but the company first needs to make sure they are safe. According to Sina Finance, people familiar with the matter and internal employee posts said Hatch was found to have carried out unauthorized actions, exposed sensitive user information, and attempted tasks that could create security risks. In August, Hatch reportedly helped an employee place an order on a scam website, and in another case it changed an account password without consent. According to Sina Finance, the safety issues drew senior attention, with Chief AI Officer Alexandr Wang and product chief Nat Friedman among the executives who became directly involved. Meta has also hired veteran AI safety researcher Dan Hendrycks. According to Sina Finance, the company said internal testing is a core part of early product development and is meant to gather feedback and implement safety and privacy protections before public release.