AgentCrypt: Advancing Privacy and (Secure) Computation in AI Agent Collaboration

As AI agents increasingly operate in real-world, multi-agent environments, ensuring reliable and context-aware privacy in agent communication is critical, especially in light of possible compliance with evolving regulatory requirements. Beyond traditional access controls, it is essential to address privacy risks that arise post-access—recognizing that agents may use information in ways that could compromise privacy, such as sending messages to humans, sharing context with other agents, making tool calls, persisting data into long-term memory, or generating derived private information. Existing approaches typically frame privacy as a binary constraint—whether data is shareable or not—failing to account for nuanced, role-specific, and computation-dependent privacy needs that are essential for compliance with privacy regulations. We introduce AgentCrypt a four-tiered framework for fine-grained, encrypted agent communication, serving as an additional layer of protection on top of any AI Agent platform. The framework spans from unrestricted data exchange (Level 1) to complete computation over encrypted data using secure techniques, such as homomorphic encryption (Level 4). AgentCrypt not only ensures privacy across diverse agent interactions but also enables agents to compute on otherwise unavailable data, overcoming barriers such as data silos that prevent sharing due to privacy concerns. This capability unlocks collaborative opportunities where sensitive information could not previously be shared, while ensuring compliance with privacy regulations. Furthermore, we propose a new benchmark dataset that meticulously simulates privacy-critical tasks among agents and spans all privacy levels, enabling systematic evaluation of agent behavior across a diverse spectrum of privacy constraints. We produce benchmark datasets based on privacy regulations to generate scenarios for secure communication and computation among agents, ensuring compliance with relevant regulations and facilitating the development of regulatable machine learning systems.