Pwn2Own Berlin is testing Codex, Claude Code and Cursor. For teams adopting coding agents, the lesson is clear: treat the agent runtime as privileged software, not just a smarter editor.
AI app builders make internal tools and prototypes visible very quickly. The real risk appears when nobody knows which apps exist, which data they touch and who owns them.
Persistent AI memory can make assistants more useful, but it also creates a new attack surface. Companies need clear rules for what agents remember, retrieve and trust.
Tools like Lovable, Replit and similar platforms make web apps visible quickly. Before customer data, real users or internal workflows are involved, teams need a clear production and security handover.
Once AI agents can use tools, APIs and internal systems, authentication is not enough. Companies need clear execution rights, approvals and auditability.
AI agents can accelerate software delivery. In production workflows, they also need scoped permissions, reviews, tests, logs, and operating rules.