Early accessSecurity

Security and data handling

How Humane isolates tenants, authenticates calls, signs what it sends and removes data when asked. This page sticks to controls that exist in the product today.

Python SDK · JavaScript SDK · REST API · Any LLM

Workspace · your team

Members

  • Anaowner
  • Rajadmin
  • Samviewer

API keys

  • hx_…a1f3active
  • hx_…7c20rotated
  • hx_…e914revoked
Google sign-inKeys stored hashedTenant-scoped data
Illustration of a team workspace
Controls

What's in place

Per-tenant isolation

Every stored row is scoped to a tenant and cross-tenant reads are blocked in the query layer. Each tenant's memory lives in its own store.

API-key authentication

hx_ keys are stored as SHA-256 hashes and matched exactly. They work only on SDK routes, so a leaked key can't create keys, invite admins or read billing.

HMAC-signed webhooks

Every delivery is signed with HMAC-SHA256 over a timestamp and the raw body. The SDK verifiers reject bad signatures and anything older than five minutes.

SSRF guards

Every URL Humane will call for you, from webhook targets to model base URLs, must be http(s) and must not resolve to loopback, private, link-local or cloud metadata addresses.

Export and erasure

One call exports what Humane holds on a user. Erasure removes their database records and vector memory, and a failed memory delete is reported, not hidden.

Retention

Stored conversation messages are capped per plan, enforced by an hourly sweep that runs off the request path.

Personal data in replies

Third-party identifiers are masked in outbound replies, and SSNs, card numbers and IBANs block the reply.

Audit with redaction

Audit logs and exports redact personal data for non-admin roles. Admin actions such as key revocation are recorded.

Abuse limits

Rate limits per API key, a stricter limit on model-backed drafts, and login throttling that can't be used to lock someone out of their own account.

Both directions

Signed on the way out, verified on the way in

Webhooks Humane sends carry a signature, an event id and a delivery attempt number. Messages arriving from channels are checked too: Telegram updates must carry the channel's secret token and WhatsApp webhooks must carry a valid app-secret signature.

  • Retries with a dead-letter queue and manual retry
  • Channel secrets shown once, then redacted to the last four characters
  • Event names validated at subscription time
  • Webhook and model URLs checked against SSRF rules before every call

Webhook delivery

POST /hooks/humane
X-Humane-Event: safety.flagged
X-Humane-Delivery-Attempt: 1
X-Humane-Signature: t=1762000000,v1=5f2c…9ab1

{"event":"safety.flagged",
 "data":{"user_id":"user_88","risk_level":"critical",
         "flags":["self_harm"],"source":"observe"}}
HMAC-SHA256(secret, t + "." + body)Valid

Retried on failure · dead-letter queue · manual retry

Illustration of a signed webhook delivery
Hardening

Details that came out of our own red-team review

  • Production refuses to start with safety-bypass or debug flags set
  • Malformed input returns a clean 4xx instead of a server error
  • Password reset gives the same answer whether or not an account exists
  • Provider error messages are cleaned before they reach users
  • Verified email required before an API key can be created
  • Owner or admin role required to change engines
Data rights

Export and erase a user's data

python
dump = client.export_user_data("user_42")   # everything Humane holds on this user
client.erase_user_data("user_42")           # records, history, tasks and vector memory

Certifications

Humane doesn't hold third-party certifications such as SOC 2 yet. If your review needs specific controls or documents, tell us early and we'll walk your security team through the product.

Contact us

Security questions

Something else? Ask us.

Questions from your security team?

We're happy to go through the controls in detail as part of early access.