Securing Agentic Identity (codon.org.uk)

A recent discussion in the security industry highlights the growing challenges posed by the integration of large language models (LLMs) into enterprise environments, particularly concerning the management of sensitive credentials. As organizations increasingly empower LLMs to interact with various resources, such as calendars and emails, the risk of information leaks rises when these non-deterministic agents handle access tokens. Traditional security mechanisms are often inadequate, leaving organizations vulnerable to credential theft and unauthorized access, especially when access tokens and sensitive data are improperly stored or managed. To address these security concerns, a new idea has emerged involving a broker service that centralizes credential management while enhancing security. This service allows agents to request tokens through a secure proxy without exposing real tokens in the agent environment. Instead, the broker generates a signed JWT containing an encrypted version of the token. This approach limits the exposure of sensitive credentials, relies on mutual TLS for verification, and ensures that tokens are only usable in intended environments. Although these improvements still face challenges related to third-party integrations and opaque tokens, they signify a step toward more robust security frameworks that can better protect enterprise systems relying on AI/ML technologies.