Security

20+ detectors across 5 categories — contact, identity, financial, secrets, network — scan the messages exchanged with your agent.

Each guardrail can mask the detected data (replaced with a placeholder like [REDACTED_EMAIL]) or block the request outright, depending on what you configure per API key.

• —IP whitelist: each key only accepts traffic from the addresses you declare.
• —Model whitelist: a compromised key can't switch to a more expensive model than the ones you've allowed.
• —Kill switch and time windows: cut off a key's access in one click, or restrict it to a specific time window.

A requests-per-minute budget protects the shared infrastructure, both per tenant and per key. A concurrent-call ceiling bounds what an integration can actually hold in memory at any moment, independent of its throughput.

Your provider keys (OpenAI, Anthropic, etc.) are encrypted at rest with AES-256-GCM. They are never readable in plain text, not even by us.

Security and access control always apply, even in degraded mode: only optimization is removable, never the protection of your data.