Security & compliance

Trust & Security

Enterprise-grade by design — how every conversation, action and token stays safe. Each claim is tagged so you can see what is live in this build versus planned or via a provider. Live-LLM hero-set rates and connector proofs live on the Live quality scoreboard.

POPIA-alignedGDPR-readyPCI-DSS (via provider)SOC 2 (in progress)

Live — in this buildPartial — real capability, fuller story on roadmapVia provider — Stripe / host / IdPPlanned — not claiming today

Data & privacy

Tenant data isolationLive
Every business is workspace-partitioned — audit and APIs never return cross-tenant rows.
Encryption in transit & at restLive
TLS on the host edge · OAuth tokens encrypted with AES-256-GCM at rest (v2 envelopes).
Legacy v1 HMAC seals still open for migration; Azure Key Vault wrapping is next.
PII minimisationPartial
Agents collect only what the task needs; connector secrets never enter prompts. Turn transcripts are redacted at write time (Phase 3).
Best-effort redaction (email/phone/card/OTP) — not a substitute for data minimisation at source.
Retention & erasureLive
DSAR JSON export and admin-managed erasure API for owners/admins. Audit trail stays append-only with tombstone events.
Erasure is destructive and admin-only — chat agents still escalate to a human for end-user requests.

Identity & access

OAuth 2.0 + PKCELive
Live in this build — every connector authorises with PKCE and HMAC-signed state.
SSO / OIDCPartial
OIDC Bearer adapter ready for MyInstantAI identity (`workspace_id`, `roles`).
SAML via your IdP once MIAI SSO is wired — mock auth is for demos only.
Role-based accessLive
Owner · admin · agent · read-only enforced on rent, wallet, configure, knowledge, OAuth, DSAR. Workspace page manages members and roles.
OIDC still trusts IdP roles until SCIM; in-app invites are token-based (no email send yet).
Scoped embed keysLive
Per workspace + agent, HMAC-bound, rate-limited — only live/rented agents serve chat.

AI safety & guardrails

Grounded answersLive
Fenced to knowledge + tools — agents must not invent policy, prices, or records.
Prompt-injection defenseLive
Guardrail docs + runtime input/output checks in Agent Studio chat.
Human-in-the-loopLive
Sensitive, disputed, or high-value actions hand off to a person — including erasure requests.
Output moderationPartial
Off-scope and unsafe replies are blocked or rephrased by pack guardrails before send.
Policy-layer moderation in runtime packs — not a separate third-party moderator yet.
AI transparency (Art.50-aligned)Live
Chat, embed, and App surfaces disclose that replies come from an AI system — not a human — with a clear path to request a person.
Draft disclosure copy pending counsel review for jurisdiction-specific wording.

Infrastructure & ops

Secrets stay server-sideLive
Connector credentials encrypted at rest and never shipped to the browser.
Azure Key Vault is on the deploy path; staging uses env + encrypted file store.
Full audit trailLive
Rent, chat, embed, OAuth, knowledge, wallet, consent, and DSAR export/erasure events per workspace.
Append-oriented Postgres inserts; true WORM / immutable blob export planned with SOC 2 evidence.
Rate limits & spend capsPartial
Embed: 30 req/min per key · wallet pause when tokens run out.
Richer per-tenant quota UI and anomaly alerts are next.
Security headers + CSPLive
Enforcing Content-Security-Policy plus nosniff, frame, referrer, and permissions policies.

See it live

Open any agent's studio chat and try an unsafe prompt — the runtime blocks prompt-injection, refuses cross-tenant data, and escalates erasure requests to a human.

Open AI Agents

Regional packs

Agent-layer compliance by market. Not the same as platform certification or residency guarantees.

United States

CCPATCPAHIPAA (health — BAA required)

US packs bake TCPA (no cold outreach, honor STOP), CCPA minimisation, and health-agent HIPAA caution into prompts and guardrails.

European Union

GDPRePrivacy (marketing)

EU packs minimise personal data and hand access/erasure requests to a human — DSAR export for portability.

Africa (incl. South Africa)

AFRICA

POPIA-styleRegional privacy

Africa packs follow POPIA-style collection limits and escalate sensitive HR/payroll/medical matters to a person.

Asia

ASIA

PDPA-styleRegional privacy

Asia packs apply PDPA-style minimisation, opt-out via human handoff, and no cold marketing messages. Distinct from Oceania (AU/NZ/Pacific).

Oceania (AU / NZ / Pacific)

OCEANIA

Australian Privacy Act (APPs)NZ Privacy Act

Oceania packs bake AU/NZ privacy minimisation, STOP handling, and dual emergency guidance (000 / 111) into prompts and guardrails. Default currency AUD with NZD called out for NZ tenants.

Workspace audit & DSAR

This workspace only — never cross-tenant. Owners/admins can export a DSAR JSON pack or request destructive erasure (admin-only).

Loading…

Roadmap

Now
RBAC matrix, AES-GCM tokens, DSAR export, CSP enforce, runtime guardrail refusals
Shipped
Next
Key Vault wrapping, CSP nonce tightening, invite email delivery / SCIM, anomaly alerts
In progress
Azure cutover
Per-tenant region pin, EU/ZA/APAC residency options, App Insights SIEM hooks
Planned
Post-GA
SOC 2 Type II, DPA + subprocessor schedule, breach SLA
Planned

Full write-up: docs/TRUST_AND_COMPLIANCE.md. Partner security: security@myinstantai.com.

Loading…

Trust & Security

POPIA-alignedGDPR-readyPCI-DSS (via provider)SOC 2 (in progress)

Live — in this buildPartial — real capability, fuller story on roadmapVia provider — Stripe / host / IdPPlanned — not claiming today