AI agents that handle third-party API tokens are vulnerable to credential theft because they process untrusted input, making it necessary to adopt proxy-based authentication to secure user data.

Key Points

• AI agents often leak OAuth tokens through logs, error reports, context windows, and tool call arguments during normal operation.
• Prompt injection attacks can trick agents into exfiltrating these tokens to unauthorized third-party servers.
• Traditional mitigations like scope narrowing and token rotation fail to address the fundamental risk of storing credentials in agent runtimes.
• Credential proxying services, such as WorkOS Relay, allow agents to perform actions without ever holding the actual access tokens.
• Proxying restricts agent capabilities to allowlisted hosts and prevents the permanent theft of bearer credentials.

Why it Matters

Moving sensitive credentials out of the agent runtime significantly reduces the blast radius of a potential security breach. By centralizing authentication through a proxy, organizations can ensure that even a compromised agent cannot permanently exfiltrate user access tokens.