Kim Maida uses an on-call triage demonstration to show how standing API keys and database credentials let an agent expose customer data, drop a production database, restart services or incur new infrastructure costs. Private data, untrusted inputs and external communication create a dangerous combination when useful capabilities are paired with excessive standing access. Human approval alone can shift the security burden to a tired user.
Kim Maida explains a token-exchange architecture that records the user an agent acts for, the requesting application and the delegated permissions. A security token service evaluates policy before issuing ephemeral credentials for a specific tool call, with further exchanges for downstream resources. In the revised demonstration, policy blocks customer-data access and destructive database operations, permits certificate renewal, and prevents an unauthorized restart even after user approval.
Kim Maida distinguishes this demonstrated policy-governance flow from Keycard's proposed intent-based authorization approach, which adds an approved mission with permitted actions and limits. She explicitly says intent-based authorization is not a finalized standard and that the demonstration does not implement it. The lesson is to enforce access through identity, scoped credentials and policy, rather than treating prompts or consent as sufficient security boundaries.
Watch on YouTube
