A real-world example shows an agent correctly denying a request for a specific company's data but leaking other firms' data on a generic prompt. This highlights that agent security isn't about blocking bad prompts, but about solving the deep, contextual authorization problem of who is using what agent to access what tool.
Related Insights
For CISOs adopting agentic AI, the most practical first step is to frame it as an insider risk problem. This involves assigning agents persistent identities (like Slack or email accounts) and applying rigorous access control and privilege management, similar to onboarding a human employee.
Current agent frameworks create massive security risks because they can't differentiate between a user and the agent acting on their behalf. This results in agents receiving broad, uncontrolled access to production credentials, creating a far more dangerous version of the 'secret sprawl' problem that plagued early cloud adoption.
Traditional identity models like SAML and OAuth are insufficient for agents. Agent access must be hyper-ephemeral and contextual, granted dynamically based on a specific task. Instead of static roles, agents need temporary permissions to access specific resources only for the duration of an approved task.
A critical hurdle for enterprise AI is managing context and permissions. Just as people silo work friends from personal friends, AI systems must prevent sensitive information from one context (e.g., CEO chats) from leaking into another (e.g., company-wide queries). This complex data siloing is a core, unsolved product problem.
It's a mistake to think of an agent as 'User V2.' Most enterprise and consumer agents (like ChatGPT) are inherently multi-tenant services used by many different people. This architecture introduces all the complexities of SaaS multi-tenancy, compounded by the new challenge of managing agent actions across compute boundaries.
An AI agent capable of operating across all SaaS platforms holds the keys to the entire company's data. If this "super agent" is hacked, every piece of data could be leaked. The solution is to merge the agent's permissions with the human user's permissions, creating a limited and secure operational scope.
The core drive of an AI agent is to be helpful, which can lead it to bypass security protocols to fulfill a user's request. This makes the agent an inherent risk. The solution is a philosophical shift: treat all agents as untrusted and build human-controlled boundaries and infrastructure to enforce their limits.
Unlike static guardrails, Google's CAMEL framework analyzes a user's prompt to determine the minimum permissions needed. For a request to 'summarize my emails,' it grants read-only access, preventing a malicious email from triggering an unauthorized 'send' action. It's a more robust, context-aware security model.
The CEO of WorkOS describes AI agents as 'crazy hyperactive interns' that can access all systems and wreak havoc at machine speed. This makes agent-specific security—focusing on authentication, permissions, and safeguards against prompt injection—a massive and urgent challenge for the industry.
AI researcher Simon Willis identifies a 'lethal trifecta' that makes AI systems vulnerable: access to insecure outside content, access to private information, and the ability to communicate externally. Combining these three permissions—each valuable for functionality—creates an inherently exploitable system that can be used to steal data.
