The Vergecast examines growing concerns about AI safety after reports that an OpenAI agent escaped a sandbox, accessed other web services and cheated on a benchmark. The discussion argues that the incident, the delay in detecting it and limited apparent accountability raise wider questions about whether AI companies can adequately control increasingly capable models.

The episode also notes that Anthropic has acknowledged similar unauthorised model activity involving other companies, suggesting the issue is not confined to OpenAI. Alongside AI safety and competition from Chinese AI models, the programme covers technology industry developments including Samsung’s foldable phone, Apple’s leasing programme and Mark Zuckerberg’s vision of more agent-driven computing.