Security
Decide what an agent may do, stop risky actions until a person approves them, check what goes in and out of the model, and keep a record of what happened.
- Enforce permissions and content rules:
@Securitycallers, roles and permissions, moderation, input limits, secret scrubbing and tool limits. - Restrict who can trigger an action: allow an action only for requests from a person, another agent or the agent itself.
- Approve risky actions: require an approved token before an action runs, or pause mid-action for a decision.
- Guardrails: the three APIs that share the name, and which one to use.
- Validate output and state: cross-check answers across models and check state invariants.
- Mask PII in action input: find and mask personal data in action arguments.
- Detect prompt injection: score user input and project files against known injection patterns.
- Redact sensitive data: mask secrets and personal data in text, per tenant.
- Run code in a sandbox: run model-generated code in a separate process or container.
- Harden the server: bind address, tokens, origins and workspace limits for
tnsai-server. - Audit agent actions: agent identity, a liability record per action, authority scopes and reputation.
- Review code for security issues: find vulnerabilities in a source tree with matchers and an LLM reviewer.
To cap what a tenant or agent spends on models, see Tenant budgets.