Two Anthropic leaders explain in a 107-minute interview how they keep Claude's AI agents safe, including how passwords are protected. Useful for any business weighing agents.
Two Anthropic leaders, Katelyn Lesse and Angela Jiang, have explained how the company tries to keep its AI agents safe. The talk ran 107 minutes on the Latent Space podcast. Their core idea is to keep the model's thinking apart from the tools it uses to act. They call these the "brain" and the "hands."
An AI agent is software that does tasks for you, not just answers questions. The interview covers the Claude Platform, the launch of Claude Opus 5.5, and how to build managed agents that stay reliable. The chapter list also points to protecting credentials, which means passwords and access keys, and to keeping agents in isolated sandboxes. A sandbox is a walled-off space where a program can run without touching everything else.
Other topics include safety classifiers, model routing, and rebuilding the Messages API in Rust. The show notes link to Anthropic pieces on isolation, credentials and network controls.
The material is a description and a list of chapters, not a transcript. It does not say exactly how the safeguards work or how well they perform. Anyone planning to use agents should ask a provider how it protects passwords and limits what an agent can reach. Those two questions are the practical lesson here.
Before using an AI agent at work, ask the provider how it protects your passwords and limits what the agent can touch.
A new Alibaba-led test of real coding requests found the best AI agent fully finished just 26.4 percent of tasks, so human review still matters.
Read more: AI coding helpers fully finish only about 1 in 4 real tasksEA has re-released 25 classic PC games at a discount until November 19, and Bethesda has finally given The Elder Scrolls VI a subtitle: Sentinel.
Read more: EA brings 25 classic PC games back, and Elder Scrolls VI gets a nameA user says Anthropic's Claude Code put his email in web requests, then scam messages arrived. The link is unproven, but it shows what AI tools can share.
Read more: User says Claude Code sent his email address to websites
