CompanyAten Security joins San Francisco's Emerging Technologies poolRead the post →

Security and trust

A security review, in the order your reviewers ask.

Short answers first. Each points to the setting, the document or the record that backs it. Where we can’t show evidence yet, we say so.

Hosted
In the US
Azure West US 2 and AWS us-west-2
When it fails
Held, not waved through
If your policies cannot be loaded
Decision records
90 days in the product
Stream them to your SIEM to keep them longer
Evidence
Checked offline
Signed packs, verified with thothctl
Assurance
SOC 2 Type I, ISO 42001
Evidence linked below

01 · Assurance

What we hold, and the evidence for it.

Each badge opens the evidence. We show nothing we can’t link.

Select a mark to see what it is and where the evidence is.

In preparationSOC 2 Type IIGDPRCISA Secure by Design pledgeNo badge until each is complete.

AARM · runtime implementer in the technical working group

AIUC-1 · mapping our decision records, in progress. No conformance claim.

02 · Your data

What we keep, what we don’t, and for how long.

You choose, per workspace, whether tool arguments are stored, hashed, redacted or dropped. A change applies to data that arrives after it.

DataDefaultKept for
Tool argumentsHashed, with a key per workspaceWith the decision record
Prompts, model output, tool resultsDroppedNot kept
Decision recordsStored90 days
What an approver sees on a held actionMasked, at most 500 charactersDeleted when the action is resolved
Ask Thoth conversationsVisible only to the person who started them90 days after the last message
A trial workspaceThe same defaults30 days, then 30 days read-only, then deleted

Workspace settings

Data handling

Tool arguments

Argument names are kept. Each value is kept as a hash, not readable.

Prompts, model output and tool results

Nothing is kept.

Defaults shown: arguments hashed, prompts and results dropped. You change them per workspace.

Tenant boundaries

Tenant-scoped authorization and storage boundaries are designed to prevent one customer from accessing another customer’s data.

Retention, exports and deletion

Confirmed for the deployment you select and your agreement, not assumed.

Read how data handling works

03 · Access

Who can see and change what?

Administrative and service access is scoped by role, and an integration only gets the permissions for its stated purpose.

  • Changes need an admin and a recent second factor. Every change is previewed first, then confirmed in the browser. Viewers read; they cannot change anything.
  • Approvers see a masked preview, not the call. Credentials, emails and phone numbers are masked, and values under secret-looking names are hidden. With hash or redact, only argument names show. With drop, nothing does.
  • Receipts and evidence hold a hash, not the arguments. A reviewer can prove what was decided without the data being kept.
  • Slack reviewers are members of your organization. Approving from Slack needs a verified email that matches a member.
  • Control who signs in. Org admins can require MFA, set how long a sign-in lasts and limit sign-in to an IP allowlist.
  • Keep secrets out of the record. Secrets go through supported secret-management paths and stay out of policy content and decision evidence.
  • You keep your side. Source identities, tool credentials and resource permissions stay with you, and so does how your applications respond to decisions.

04 · When something fails

What happens when something fails?

Failure behavior belongs in policy. A sensitive workflow can stop or ask for approval, while a lower-risk one can use a fallback you define.

If Thoth can’t be reached, the call is refused.
In enforce mode, the SDKs refuse the call unless you choose to fail open yourself.
A model can raise a decision, never lower it.
A risk score can only make a decision stricter, so no model produces an allow.
Watch first, then enforce.
Observe mode records what enforcing would have decided and changes nothing.

05 · Evidence

How do you prove it later?

Each decision keeps integrity data with its record, so later changes can be detected within the evidence system you configure.

  • The decision and the reason for it
  • The policy and rules that applied
  • A timestamp and integrity data
  • Export references for downstream review

Build a signed evidence pack for a period of up to 31 days and verify it offline with thothctl evidence verify: no network, no account, no call to Aten. Receipts and packs are signed with separate keys, and the fingerprints are published.

What verify checks

  • Every part is present and in order
  • Each part matches its digest
  • The manifest, attestation and receipt signatures
  • The event chain: hashes, signatures and continuity

Today you build and verify with thothctl, the API, Terraform or MCP. The dashboard can’t download or verify a pack yet.

See the signing keys

06 · Where it runs

Hosted in the US today. Other options, said plainly.

Available

Hosted by Aten

Compute in Azure West US 2. Queues, keys and storage in AWS us-west-2. Customers are separated by workspace.

Upon request

A dedicated stack, or your own cloud account

Tell us your requirement and we will say what is possible.

Coming soon

Self-hosted or air-gapped

Thoth is hosted only today. We will say so plainly until this is available.

07 · Subprocessors

Who else touches the data.

ProviderWhat it does for Thoth
Microsoft AzureHosting, databases and backups
Amazon Web ServicesEmail delivery, key management, secrets, evidence and report storage
WorkOSSign-in
Azure OpenAIPowers Ask Thoth
StripeBilling on paid plans
SlackApprovals and notifications, when you connect it

08 · Evaluate it

Try it without touching production.

The path a security team takes before a pilot, in four steps.

Select a step to follow the path.

By industry

What your reviewers will look for.

Third-party risk

Assurance, subprocessors, hosting location and retention are on this page, each with its source.

Audit evidence

Signed evidence packs your auditor verifies offline. Stream decisions to your SIEM to keep them as long as your policy requires.

Change control

Every change is previewed, then confirmed by an admin with a recent second factor.

Thoth sees what your integration sends. Your tools stay yours.

The integration passes the fields you configure for evaluation: identity, task, session, tool, resource and environment. Customer tools and resources stay outside Thoth’s control plane.

Step 1 of 4

Your agent or application

Your code runs the agent and holds its credentials. Before a tool runs, your integration asks Thoth what to do and waits for the answer.

Source identities, tool credentials and resource permissions stay with you.

Select a step to see what it does and what crosses it.

Questions risk teams ask

The short answers.

Which models review our data?

Only US-origin models: OpenAI models on Azure OpenAI, or OpenAI’s open-weight models on an OpenAI-compatible endpoint.

What does Thoth not decide?

It does not replace your approval process. It holds the action and shows a person what it would do, and a person decides. It does not claim to make an agent safe: it checks each action against your policies before it runs, and keeps the record.

Do you hold certifications?

SOC 2 Type I and ISO/IEC 42001, with the evidence in the Trust Center. SOC 2 Type II, GDPR and the CISA Secure by Design pledge are in progress, and we show no badge for any of them until it is complete. AIUC-1 is a mapping in progress, not a conformance claim.

How do I report a vulnerability?

Email security@atensecurity.com. The disclosure page lists scope and safe-testing guidance. We ask you to coordinate before publishing.

Where we are today, said plainly.

In place

  • Security program. Documented practices, with product controls reviewed as the platform changes.
  • Responsible disclosure. Researchers can report a vulnerability directly to us for coordinated review.
  • Customer diligence. Current security documents and available independent evidence are available through the Trust Center.

What we haven’t claimed

  • Framework support. Product controls are designed to support your mappings to common security and AI-risk frameworks. What applies depends on your deployment and configuration, and we’ll publish verified assurance updates through the Trust Center.

Backed by

Advised by security leaders from

In standards work with

Mapping our decision records to

Review your security requirements with Aten.

Bring your questionnaire. We’ll walk through boundaries, data handling, failure behavior and the assurance material available for your evaluation.

For technology, channel, federal delivery, and research partnerships.

Help improve Thoth. Allow Amplitude and Mixpanel to measure page use, interactions and errors across the Aten Security website, documentation, signup and dashboard. We exclude customer content and credentials. Your choice does not affect access. You can change it in Analytics preferences.