Security

How Halto handles your data, and where it stops

Halto sits between AI agents and a business's systems, so it sees what passes between them. This page says what that is, what Halto keeps and for how long, and what it does not protect against. Found a vulnerability? Write to team@halto.tech; we answer within 3 working days. Please tell us before you tell anyone else.

What passes through Halto

An agent sends Halto the action it wants to take (tool, action, parameters) and, with the data guard, the text it is about to read. Halto answers allow, pending or block, and replaces personal and confidential values in that text with placeholders such as [[EMAIL_K7Q2M]] before the agent's model reads it. The real values go back into an action only where your rules let them go.

Halto itself sends your data to no AI company: the model that finds names and addresses runs on Halto's own server in Frankfurt.

The model gateway forwards your own requests to the model you chose, under your own key, with values already placeholders. That is your request passing through, not Halto sending data.

The privacy policy lists every service that processes data for Halto.

What Halto keeps, and for how long

Decisions
Each decision with its reason, the rule that applied and who approved it, for as long as you choose, up to 12 months.
Hidden values
Encrypted with your workspace's own key, for 24 hours by default (1 hour to 7 days, your choice), so they can be put back.
Rules and settings
With a history of who changed what.
Never in the log
Passwords, API keys, card numbers and IBANs are removed before anything is written to the log or an error report. The text of conversations through the model gateway is not stored.
Card numbers
Never in the log. A card the data guard hides is kept only in the encrypted store above, so it can be put back, until it expires.
Dashboard use
Which pages and buttons people use, never what they type: 90 days, seen only by the people who run Halto.
Never done
Halto never holds, moves or issues money.

Deleting your data

Settings has a download of everything Halto holds about your workspace, and a button that deletes it for good.

Deleting a workspace removes its data from Halto's database at once. Decisions past your retention period are deleted automatically.

Where it runs

Region
Frankfurt, Germany (EU): the application, the database and Halto's model.
In transit
HTTPS only, with HSTS.
At rest
The database is encrypted at rest by its provider. Agent keys are stored only as hashes. Secrets a workspace stores (tool credentials, tokens) are encrypted again with that workspace's own data key.
Isolation
Row-level security on every table, so one workspace's queries can't read another's.

A log you can check yourself

Every 5 minutes Halto seals the new decisions and approval answers: each seal hashes its records and the seal before it, signed with Ed25519. The public key is at /.well-known/halto-log-key.

A record is sealed once it is 5 minutes old, so the newest records, up to about 10 minutes, are not sealed yet.

Download the log from Reports → Signed log and check it at /verify-log.

The latest fingerprint also goes into the weekly email owners get on Mondays and into the oversight report (the PDF under Reports), so you hold a copy: not even Halto can rewrite history without it showing.

Who can approve and change things

  • Sign-in confirms the email address; once someone has Face ID or a fingerprint, every new sign-in needs it (or a single-use recovery code).
  • Owners and admins must set up Face ID within 7 days; after that, changing rules, keys, the team or settings needs it.
  • Loosening protection (releasing the emergency stop, resuming a paused agent, new keys, new approvers) needs a second factor every time.
  • Only a linked approver can approve: a link in an email never approves by itself, and risky requests ask for Face ID.
  • Halto's staff: the two people who run Halto can reach production, through accounts (GitHub, Vercel, Supabase, Google) with two-factor sign-in.
  • They hold the server secret that protects workspace keys, so they could decrypt hidden values. Their direct database access isn't logged per person yet.
  • Halto's admin screen shows accounts, plans and counts of actions, never what the decisions, rules or texts were.

If Halto is down or can't decide

The action doesn't run. The SDKs throw instead of returning a decision, the no-code step fails and the workflow stops, Halto Guard doesn't run the command, and tools behind the MCP gateway can't be reached. An error inside Halto blocks the action too. There is no setting that lets actions through instead.

Measured

Decision time, real traffic
14 ms median inside Halto: the 2,023 checks on the live service in the 30 days to 13 November 2026, the load test left out.
Decision time, load test
110 ms median: 1,169 checks sent at once on 12 November 2026. Over all 3,192 checks the median is 19 ms.
Network
Time to reach Halto comes on top of these. Today's figure is on the status page.
Tests
3,489 tests of the rule engine and 1,309 of the app run on every change, including property tests that generate names, wrappings and parameters instead of listing known ones.
Security reviews
Ten rounds of code review so far, one from outside Halto and the rest our own. Every confirmed finding is fixed and kept as a test.
Certifications
None yet: Halto has no SOC 2 or ISO 27001 report and has had no penetration test by an outside firm.

What Halto does not protect against

  • What doesn't pass through it. An agent with its own direct credentials to a tool can act without asking. The strongest setups give agents credentials only to Halto: the MCP gateway, the model gateway, or Halto Guard on the computer.
  • An action described falsely. Through the SDK or API, Halto judges the action as the agent describes it. Behind the MCP gateway or Guard, Halto is the path itself, so there is nothing to describe.
  • Every possible prompt injection. No check finds every hidden instruction. Rules limit the damage when an attack gets through: payments, exports and new recipients can wait for a person whatever the agent read.
  • Every name and address. The built-in rules miss some; Halto's model finds more but is slower. Images and scanned documents pass unread.
  • Reading in transit. Through the model gateway, a request's text is readable on Halto's servers while it runs, though never stored. A version for your own servers isn't built yet.
  • Someone with admin rights on a computer. They can remove Halto Guard. A company install reports a computer that stops checking in.
  • Every AI app. The browser extension is in testing and covers ChatGPT, Claude, Gemini and Copilot in Chrome and Edge; phone apps and other browsers are covered only if the company's own device policy blocks them.