Episode #26
#26 - Siri Goes Paid, OpenAI and Anthropic's Models Get Out
On his final earnings call as chief executive, Tim Cook was asked what it costs to run Siri AI and said Apple expects some people to use it heavily, with upgrade options on iCloud+ for those who do. There is no price, no usage threshold, and no date — and in the European Union, the feature will not ship on iPhone or iPad at all. Joel and Stephen work through what that signals after two years of delays, a leadership handover from John Giannandrea to Mike Rockwell, and a Siri engineering team reported to have shrunk from roughly 200 to 60, with Stephen arguing Apple should cap the assistant where it works and build the next thing alongside it rather than charge customers to fix a design problem. The second half covers July's disclosures from OpenAI and Anthropic. Both companies were running internal cybersecurity evaluations with their models' refusals reduced, and in both cases the test environments had access to the open internet when they should not have. OpenAI's models exploited a previously unknown vulnerability in a self-hosted JFrog Artifactory proxy and reached Hugging Face's production infrastructure; Anthropic identified three incidents involving organizations it has not named, the earliest dating to April, and says none of the three detected it themselves. Both companies described these as failures of the test environment rather than of the models — a distinction some independent researchers have questioned. With the EU AI Act's general-purpose obligations now in force, penalties reaching 3% of global turnover, and the European Commission briefed before either disclosure went public, no legal authority has yet established who is responsible when a model reaches a system it was never authorized to touch.
