WeaknessASC-2026-0003Major weakness
Identity and privilege abuse by agents
HighMitigated
Agents inherit a human's or service's broad credentials, or reuse long-lived tokens across tasks and users.
How to detect it
- MCP server or tool handles tokens, secrets or passwords directly
- Server requests broad OAuth scopes or static API keys
- Secrets appear in agent output or tool arguments
How to resolve it
- Issue short-lived, task-scoped credentials to the agent's own identity
- Never pass user tokens through to downstream tools unchanged
- Log which identity performed every tool call
How to investigate
- Audit which identity and token each tool call used, and with what scopes
- Look for token use outside the expected task, time window or network
- Revoke and rotate affected credentials, then re-issue with narrower scopes