Trust and security
Security built into every agent and every conversation
Security isn't a layer we added afterwards. Every agent is approved before it runs, every conversation is sealed in its own sandbox, sensitive actions wait for a person, state changes are recorded, and any session can be stopped.
- Sandbox per conversation
- Closed network
- No keys in the sandbox
- Human approval
- Encrypted vault
- Audit trail
Safeguards
Layered safeguards, from the first draft to every conversation
A sandbox for every conversation
Each conversation runs in its own isolated NVIDIA sandbox with CPU and memory limits. It serves that session only and is removed when the session goes idle or is stopped.
A closed network by default
Agents reach only the systems an environment allows. If an agent needs anything else, it can't be published until someone allows it.
Credentials kept out of reach
Credentials are stored encrypted and used by the platform only when a call needs them. Model keys and people's passwords never enter the sandbox.
People approve what matters
Agents are signed off before anyone can use them, and sensitive actions wait for a person. Anyone with the right to monitor can stop a session at once.
Access down to the record
Security groups and per-record permissions decide who can use each agent, Space, folder and tool. Answers only draw on what the person asking may open.
Protect and guardrails
With Protect switched on, every message is checked for prompt attacks and personal data before it reaches the model, and if the check can't run, the conversation is refused. A platform guardrail sits above every agent's instructions, and uploaded skills are scanned before use.
See it
How a published agent runs
One sandbox per conversation, and one gateway that decides what can leave.
Chapters
- 1Publish it, sandboxedSign-off, then one sandbox per chat
Our security model
The Sigma security model
- 01
Contain
Every conversation is isolated from every other. If one is manipulated, the damage stays in its own sandbox and its allowed systems.
- 02
Control
Nothing goes live without approval, nothing leaves the network without permission, and sensitive actions wait for a person.
- 03
Keep secrets secret
Credentials live in an encrypted vault and are used at the moment of the call. They are never handed to the model.
- 04
Prove it
Sessions keep their transcript and timeline, state-changing actions are recorded in an audit log, and you can explain anyone's access to anything.
Your data
Your data, your rules
Runs where you choose
Install Sigma on your own servers or in your own cloud account. The same release runs everywhere.
Answers respect permissions
Agents answer from the documents the person asking is allowed to open, and answers cite their sources.
Clear about where data goes
Model calls go to the provider you choose for each agent, and document search uses a managed index today. Local options for both are on our roadmap, and we walk you through the data flows for your setup.
Platform practices
How Sigma is built and run
- Encryption in transit for every connection to Sigma's cloud
- Credentials encrypted at rest in the platform vault
- One isolated sandbox per conversation, removed when it ends
- Network closed by default; agents reach only allowed systems
- Record-level permissions, and an explanation of anyone's access
- Audit log of state-changing actions and full session transcripts
- API keys that expire, with usage recorded per key
- Sign-in sessions that expire, under a central password policy
- Protect checks that fail closed when they can't run
- Automated security testing in our build pipeline: dynamic scans of a running install, secret scanning, code and dependency scanning
- Hardened web console with strict security headers
- Daily backups of Sigma's cloud, kept for 14 days
Responsible disclosure
Found a vulnerability in Sigma or this website? Email [email protected] with the subject "Security report", with enough detail to reproduce it. We take every report seriously, reply quickly and keep you informed while we fix it. Please give us reasonable time to fix an issue before you share it.
[email protected]Security reviews
Security reviews are a normal part of buying Sigma. We answer questionnaires and share architecture, data-flow and deployment documentation for your setup.
Make your organisation an AI organisation
Sigma is the enterprise AI platform that connects your systems, your knowledge and your people, so AI can do real work, securely. Start with one use case and grow from there.
Or email [email protected]