Latten

How do you secure AI agents that can access company data?

Start by measuring their real reach. An AI agent is only as safe as the service accounts, tools, and data domains it can use. Secure it by attributing every action to the actor behind it, mapping what it reaches from real traffic, and then bounding the wide edges after you can see them.

See it on a live graph — no signup →

Shared service accounts are the hidden risk

When many agents or users share one service account, every actor inherits that account's reach. Latten makes that inherited reach visible.

Security needs cost context too

Runaway loops and retries are not just reliability problems; they are spend and access paths with no owner. Latten shows cost and blast radius together.

Observe before enforcement

Blocking before you understand traffic creates false confidence and breakage. Latten starts observe-only, so teams can decide what to bound with evidence.

How it works

  1. 1. Instrument the agent Track model calls, tool calls, database access, and handoffs.
  2. 2. Attribute the actor Preserve who asked and which agent acted on their behalf.
  3. 3. Find broad reach Look for shared service accounts, unattributed calls, and sensitive domains.
  4. 4. Narrow deliberately Apply the smallest useful boundary and verify the finding closes.

Common questions

What is the first security question for AI agents?

Ask what the agent can actually reach through its service accounts and tools.

Does Latten copy prompts or database rows?

No. Raw values do not leave the boundary; PII is reported as types and counts only.

Can an agent install Latten itself?

Yes. The @latten/installer MCP can provision, scan, instrument, verify, and produce a receipt for a human to review.

See a live graph — no signup Instrument your app → All guides