Moving Beyond API Keys to Secure AI Agents

As AI agents become more autonomous, a new security framework called DPACT offers a blueprint for building responsible, guardrailed systems.

As AI agents grow more autonomous, the simple security models we use today are becoming dangerously outdated. In a recent InfoQ podcast, software engineer Sahil Agarwal highlighted this growing gap, arguing that static, long-lived API keys are not enough to control agents that can take independent actions. To address this, he introduced the DPACT framework, a conceptual blueprint for building safer, more responsible agentic systems. This new model shifts the focus from simple, binary access to a more nuanced system of delegated authority, providing the guardrails needed to prevent AI agents from acting outside their intended scope and causing potential harm.

At its core, the DPACT framework is an acronym for five key security principles: Delegation, Policy, Auditability, Context, and Time. Delegation means an agent should be granted specific, limited permissions for a task, rather than inheriting the full permissions of its creator. Policy involves defining clear, machine-readable rules that govern what an agent can and cannot do. Auditability ensures that every action taken by an agent is logged for review, creating a clear trail of accountability. Context requires the system to consider the situation—such as the user, device, or data involved—before authorizing an action. Finally, Time dictates that all permissions should be temporary and expire automatically, minimizing the window of opportunity for misuse. Together, these principles create a robust system that treats AI agents not as trusted insiders, but as external actors with strictly bounded capabilities.

This approach mirrors the evolution of security practices in other areas of software development. The move from a single API key to a DPACT model is analogous to the industry’s shift from basic username-and-password authentication to sophisticated, scoped access protocols like OAuth 2.0. Just as you wouldn't give a third-party application your Google password, you shouldn't give an AI agent a master key to your entire infrastructure. The DPACT framework essentially applies well-established security concepts like the Principle of Least Privilege (PoLP) and Zero Trust architecture to the unique challenges posed by autonomous AI. As agentic workflows become more common in production environments, this level of granular control is no longer a theoretical nice-to-have but a critical necessity for risk management.

For developers, CTOs, and security teams, the immediate takeaway is to begin thinking about AI agent security as a distinct and critical discipline. While DPACT is a conceptual framework rather than a specific tool, its principles can guide architectural decisions today. Teams should review how their AI systems handle identity and authorization, questioning whether a single, powerful token creates an unacceptable single point of failure. The next step is to explore implementing time-bound, scoped credentials and robust policy engines. Looking ahead, the industry will likely see the emergence of new standards and platforms that bake these principles in, making it easier to build and deploy agentic systems that are both powerful and secure.

As AI agents gain the ability to act on their own, static API keys create a massive security risk. The DPACT framework provides a much-needed mental model for engineers to build robust authorization and identity systems, preventing agents from exceeding their intended permissions and causing damage.

Deploying autonomous AI agents without proper security controls exposes companies to significant operational and data breach risks. Adopting a framework like DPACT helps mitigate these threats, enabling businesses to leverage powerful AI automation safely and build trust with customers.