AI Agents' Silent Security Crisis: 83% of Dangerous Functions Unprotected

AI agents fundamentally change the attack surface: instead of human users clicking buttons through protected UI layers, LLMs independently decide which functions to call. A scan of three major open-source codebases revealed a critical gap: 83% of dangerous functions—those that delete files, write to databases, or charge cards—have no guards whatsoever. This isn't due to negligence but invisible scope; most teams lack an inventory of their agent's capabilities and which ones are actually protected.