Dynamic Capability Scoping for Enterprise AI Agents: A Synthetic Dataset and Three-Source Permission Architecture
Enterprise AI agents are often given a "static" set of permissions, meaning they hold access to every tool they might ever need from the moment they are turned on. This creates a significant security risk: if an agent is compromised or manipulated, it can use any of these pre-granted tools, even if they are completely irrelevant to its current task. This paper proposes a shift toward "dynamic least-privilege," where an agent’s capabilities are restricted in real-time based on the specific task it is performing. By ensuring that an agent only has access to the tools necessary for its immediate job, organizations can proactively eliminate potential attack surfaces rather than just trying to detect malicious behavior after it happens.
A Three-Layer Defense Strategy
The proposed architecture uses three distinct, complementary layers to manage agent permissions:
Role-Based Ceilings: This is a hard limit based on the agent's department. For example, an Engineering agent might be allowed to query a database, while a Legal agent would be strictly blocked from doing so. This sets the maximum possible boundary for any given role.
Task-Context Classifier: This layer narrows the role-based ceiling down to the specific tools required for the current task. By analyzing the user's request, a specialized classifier predicts the minimum permissions needed. If a task only involves summarizing a document, the agent is granted only the read access required for that file, even if its role ceiling would technically allow for more.
Policy-Based Prohibitions: This acts as a final safety filter that prevents dangerous combinations of permissions. It specifically targets the "lethal trifecta" of risks: having access to private data, interacting with untrusted content, and having external communication capabilities simultaneously. If a task requires a combination of tools that violates company policy, this layer blocks the risky permission.
Building a Foundation for Evaluation
To test this architecture, the author created a synthetic dataset of 600 enterprise task prompts. These prompts cover six different departments—such as Finance, Security, and Engineering—and are mapped to a taxonomy of 15 specific tools. To ensure the dataset is robust and avoids bias, it was built using a two-pass pipeline: one pass generated the task prompts, and a separate pass labeled the required permissions. This separation ensures that the labels are based on the actual needs of the task rather than the way the prompt was written.
Refining Policy Through Simulation
A key finding from the research is that the process of creating this dataset helped improve the underlying company policies. During the initial validation, the researchers found 46 instances where a task required a tool that the department's "ceiling" did not allow. By reviewing these discrepancies, they discovered that their initial policies were often under-specified or missing realistic workflow requirements. After updating the policies to reflect these findings, the number of violations dropped by 93%. This demonstrates that generating synthetic tasks is an effective way for organizations to stress-test and refine their security policies before deploying AI agents in the real world.
Important Considerations
While this architecture provides a proactive defense, it is important to note that the task-context classifier relies on natural language, which can be susceptible to manipulation through techniques like prompt injection. Because of this, the author emphasizes that the classifier should not be the only line of defense. Instead, the system is designed as a "layered" approach where the role-based ceilings and policy-based prohibitions act as essential safeguards. The architecture is also designed to support an "observe-only" mode, which allows organizations to monitor how agents would behave under these restrictions without actually blocking them, providing valuable data for future AI safety research.
Comments (0)
to join the discussion
No comments yet
Be the first to share your thoughts!