Netskope Explains How AI Agents Amplify OAuth Token Risks

Summary

Netskope outlines how persistent, broadly scoped OAuth tokens can let AI agents execute malicious instructions without further human approval, and recommends isolated agent identities, short-lived task-scoped credentials, monitoring, and governance.

Key points

• OAuth tokens can remain valid after a user’s initial consent, allowing access without another MFA checkpoint if a token is compromised.
• Netskope estimates that 1,000 employees using five agents each could result in 50,000 active OAuth tokens, many potentially unreviewed or unrotated.
• Broad scopes, persistent credentials, limited runtime visibility, autonomous actions, and open-ended execution can increase agents’ security risks.
• A confused-deputy attack can exploit an agent’s valid token: malicious instructions in data the agent consumes may prompt it to make authorized API calls on an attacker’s behalf.
• Recommended safeguards include separate agent identities, minimal task-specific permissions, short-lived credentials, token isolation, and segmentation.
• The article also recommends agent activity logging, behavioral baselines, anomaly response, and gateways to help govern access and detect misuse.
• DPoP and Rich Authorization Requests are cited as standards that can support sender-constrained tokens and more granular permissions.

Vendors

Products

Related Articles

• GitLab Patches Critical Flaw That Could Enable Unauthenticated Server Code ExecutionAI Gateway
• Microsoft October 2026 Updates Fix 1,411 Flaws, Including Three Zero-DaysAI Gateway