Security model

Credentials

API keys and tokens live in a dedicated vault. They are injected at execution time by a tool gateway. The model never sees them: not in a prompt, not in a log, not in memory.

Prompt injection

Content fetched from email, web pages, documents and messages is treated as data, never as instructions. If a document contains text telling the agent to do something, the agent surfaces it to you rather than acting on it. High-risk tools require a human approval regardless.

Data handling

• TLS 1.2+ in transit, AES-256 at rest
• Data residency in your chosen cloud region, written into the DPA
• Your data is never used to train models. Not ours, not our providers'
• Configurable retention, down to same-day deletion

Compliance

SOC 2 Type I in progress. Annual third-party penetration tests and an open bug bounty.

Audit

Every action is logged: who, what, which system, what changed, when, and whether a human approved it. Streamable to your SIEM on Enterprise.