Secure your app: mitigate risks to agentic features

A Podkey summary of Secure your app: mitigate risks to agentic features, from WWDC 2026.Today’s big theme is AI agents getting useful enough to do real things, and risky enough that the security details really matter. The core problem is surprisingly simple: if a model can read private stuff, absorb untrusted stuff, and then go do stuff in the world, you’ve got a pretty serious attack surface. So the interesting part here is how you shrink that risk with practical guardrails like confirmations, redaction, and a very healthy suspicion of anything the model didn’t get from a trusted source.How indirect prompt injection actually worksThe Lethal TrifectaWhat counts as untrusted contextNot all tools fail the same wayThe best defenses are boring on purposeWhere the guardrails plug inRisk-based confirmations and lock-screen protectionThis podcast was created with Podkey. Make your own at https://podkey.fm