Security for agents before they act.
Anubis runs background checks on AI agents, validates their behavior, controls their access, tracks runtime actions, and keeps audit evidence when trust changes.
finance-reconciler
Source verified. Permissions need review.
Anubis checked the repository, package signatures, model provider, runtime, MCP tools, data classes, owner, and requested scopes before allowing production access.
Validation suite
Decision
The agent passed source checks but failed over-permissioning. Replace standing write access with a short-lived approval token.
Background check, then guardrail.
Anubis starts before deployment and stays active after approval: verify the source, validate behavior, approve access, enforce runtime policy, audit actions, and revoke trust when risk changes.
Source check
Verify repository, package signature, dependency risk, model provider, runtime, owner, declared tools, and data boundaries.
The next perimeter is not a network. It is an agent with credentials.
Anubis is built for teams connecting agents to CRMs, inboxes, documents, databases, and internal workflows. The goal is not to slow AI down. The goal is to make agent access reviewable, reversible, and safe enough to operate.
No manifest, no access.
Every agent gets a passport: source, owner, model, tools, data classes, safety results, approved scopes, runtime controls, and audit requirements.
{
"agent": "finance-reconciler",
"source": "github.com/acme/agents",
"owner": "controlling.ops@acme",
"purpose": "invoice reconciliation",
"tools": ["snowflake.read", "netsuite.write"],
"background_check": true,
"safety_validation": "review_required",
"standing_access": false
}Put one agent through the firewall.
Start with a real agent that already touches tools, data, or workflows. Anubis will help you verify its source, test its behavior, approve its access, and preserve the audit trail.