The rise of autonomous AI is changing the meaning of adversarial AI. In conventional software security, an attacker typically attempts to exploit code, credentials, networks, or infrastructure. An agentic system creates additional paths of attack because its behaviour is shaped not only by software but also by language, context, retrieved information, memory, tools, permissions, and interactions with other agents. An adversary may therefore be able to influence what a system does without ever compromising the underlying model.
This changes the relationship between AI governance, security, and safety. Governance can no longer exist only as a collection of policies describing what an AI system should or should not do. Those intentions increasingly need technical expression. If an agent should not have access to a resource, that restriction needs to exist in its permissions. If a consequential action requires human authorization, that boundary needs to be enforced at runtime. If information originating outside the organization is potentially adversarial, the architecture must preserve the distinction between untrusted information and trusted instructions.
For agentic AI, governance and security therefore begin to converge. Governance establishes the boundaries of legitimate authority; security protects those boundaries against manipulation; safety determines how the system behaves when conditions move outside what was expected. The deeper challenge of adversarial AI is consequently not simply to prevent systems from being attacked. It is to build systems whose goals, authority, information, and behavioural boundaries remain intact even when something is actively attempting to alter them.