Every tool call passes through a named operation, an approval for writes and a secret store that adds the key, with an audit record alongside

Security

The platform enforces the rules, whatever the prompt says.

  • Agents reach your systems only through named operations
  • No agent can approve a write
  • Keys never reach the agent or the model
  • Every action goes into a hash-chained record
8 of 11
Speaker notes

There is no general web access, so even a misled agent can only do what its operations allow. Writes need a person’s approval once per operation, unless the org leaves automatic approval on, which is the default. Credentials live in a secret store and are added on the way out. The audit record is append only and hash-chained, and maps to SOC 2, GDPR, ISO 42001, HIPAA and the EU AI Act. Personal data protection swaps personal details for placeholders before the provider sees them; it is off until an admin turns it on, and it never blocks.