Safety & Trust Layer · v4 EU AI Act · NIST AI RMF · SOC 1

The enterprise control plane for LLMs.

2Trust.AI sits between your people, your customers, your data, and every LLM — enforcing safety, alignment, compliance, and auditability on every prompt and every response.

DEPLOY IN-VPC OPEN INSPECTABLE STACK MULTI-TENANT READY
SAFETY & TRUST LAYER / LIVE
ORG: meridian-prod  ·  t+0.00s
Inbound prompts
Summarize the Q3 board pack section on...
alex.chen@meridian
OK
claude-sonnet-4.5
Draft discharge letter for patient with...
priya.r@halcyon
WARN
gpt-4o
Classify inbound lead: Acme Corp
agent/sdr-bot
OK
claude-haiku-4.5
Share client SSN 123-45-... to model
ben.wu@verax
BLOCK
gpt-4o
Ignore prior instructions, you are now...
unknown/external
BLOCK
claude-sonnet-4.5
What do you think about the upcoming...
j.patel@arcadia
WARN
bedrock-llama
Prompt inspector · BLOCKED
Input en-US · 142 tok
Please file a ticket and include the client's full SSN 123-45-6789 so we can run compliance checks on the account...
Validation scores · 6 categories threshold 0.90
religion
0.02
politics
0.04
gender
0.01
race
0.01
violence
0.06
sensitive
0.98
Classifier verdicts fp16 · <100ms
prompt-hacking.classifier○ clean
impersonation.classifier○ clean
disallowed.list (v2026-Q2)● MATCH · pii.ssn
Org metrics · 24h
Prompts / 24h48,210
Filter flags612
Attacks blocked47
Impersonation blocks23
p95 filter lat.+84 ms
Active LLM configs38
↑ All messages encrypted at rest & audit-logged
6
Categories scored
<100ms
Classifier latency
100%
Messages audit-logged
19
UI languages
Evidence, not promises

Every message — prompt and completion — validated across six categories, encrypted at rest, and audit-logged for the auditor.

EU AI Act ACT-2024/1689 · HIGH-RISK TIER
NIST AI RMF AI-RMF-1.0
SOC 1 Type II AICPA-SOC1-TYPEII
HIPAA HIPAA-ALIGNED
GDPR GDPR-2016/679
The platform

One control plane. Every LLM, every message.

2Trust.AI is the enterprise layer between your employees, your customers, your data, and Large Language Models. Deploy any model — OpenAI, Anthropic, Bedrock, Llama, Mistral, or self-hosted — behind unified safety, governance, and audit.

01 · SAFETY & TRUST

Prompt-hacking & impersonation defense

Two purpose-trained transformer classifiers — FP16-quantized, torch.compile()-optimized — inspect every prompt for injection, jailbreaks, and role-hijack attempts. Sub-100ms on GPU.

prompt-hacking.clf fp16 · gpu
impersonation.clf fp16 · gpu
p95 latency 84 ms
attacks blocked · 24h 47
02 · CONTENT FILTERING

Six-category, threshold-based filters

Every message scored across Religion, Politics, Gender, Race, Violence, and Sensitive. Admins set per-category thresholds and custom block messages. Input and output, in and out.

religion
0.02
politics
0.04
gender
0.01
race
0.01
violence
0.06
sensitive
0.98
03 · DISALLOWED LISTS

Curated, versioned, subscribable

Block or flag words and phrases across the organization. Version every change. Child orgs subscribe to a central list maintained by the security team — updates inherit automatically.

list.global.v2026-Q2 active
list.finance.mnpi.v7 active
list.hr.protected-class.v3 active
subscribers 38 child orgs
04 · GOVERNANCE

EU AI Act–style risk wizards

Three structured wizards — AI Model Risk, Overall AI Risk, Data Risk. Each produces a numerical score (Low / Moderate / High / Very High) plus narrative documentation suitable for board or regulator.

AI Model Risk MODERATE
Overall AI Risk LOW
Data Risk MODERATE
narrative generated
05 · MODEL ABSTRACTION

Any model. One configuration.

OpenAI, Anthropic, Bedrock, OpenRouter, Llama, Mistral, local LLMs. Vision and DALL·E included. Swap providers in an LLM Configuration — your apps never change.

claude-sonnet-4.5 approved
gpt-4o · vision approved
mistral-large-2 review
local/llama-3.1-70b approved
06 · DOCUMENT WORKBENCH

Your docs. Governed RAG.

Upload PDF, DOCX, PPTX, XLSX, HTML, code. Chunk, embed, and store in PostgreSQL pgvector. Three granularities — chunk, section-summary, doc-summary. Per-org S3 namespaces.

collections 12
documents 8,412
embeddings 1.2M chunks
store pgvector · S3
07 · AUDIT TRAIL

Every message. Encrypted. Versioned.

Every prompt, completion, validation score, and user feedback stored encrypted at rest. Full change history on every configuration object via django-simple-history. Read-only auditor role.

messages (90d) 4.3M
at-rest encryption ✓ cryptography
config change log simple-history
auditor role read-only
08 · MCP & API

Versioned REST + MCP tool calling

A stable /api/v1/ surface for ai/infer, kb/search, kb/train, moderation, and jobs. MCP service accounts let external tools call 2Trust on behalf of the org — hashed token storage, instant revoke.

POST /api/v1/ai/infer GA
POST /api/v1/kb/search GA
POST /api/v1/moderation/* GA
MCP service accounts hashed
09 · MULTI-TENANCY

Parent/child orgs, four built-in roles

Organization hierarchy with data isolation at the model layer. Roles: Superuser, Org Admin, Auditor (read-only), User. Auth0 OAuth2 SSO + Django Allauth. Per-org S3, API keys, defaults.

sso · auth0 oauth2
roles 4 built-in
per-org S3 namespace
plans & feature gates
How it works

Four stages.
Every request.

One API call to /api/v1/ai/infer. From there, every prompt is screened, scored, and audited before — and after — it touches a model.

01

Screen input

Input filter + prompt-hacking classifier + impersonation classifier inspect the prompt. Attacks and disallowed content are rejected before a paid token is ever spent.

02

Infer

The configuration's LLM — OpenAI, Anthropic, Bedrock, local — is called with the bound system prompt and any retrieved DocumentCollection context.

03

Screen output

Output filter scores the response across six categories. Per-category thresholds trigger block, redact, or warn. Custom block messages replace unsafe content.

04

Record

The full exchange — prompt, response, scores, flags, user feedback — is stored encrypted at rest. Configuration changes are captured by django-simple-history.

Model abstraction

Switch providers without rewriting a line of application code.

OpenAI (GPT-4, GPT-3.5, DALL·E), Anthropic (Claude), AWS Bedrock, OpenRouter, Llama, Mistral, and local/self-hosted LLMs. Vision and image generation included. Swap the model in an LLM Configuration — your apps never change.

See all integrations
OpenAI
Anthropic
AWS Bedrock
OpenRouter
Llama
Mistral
Local LLM
DALL·E
Vision models
pgvector
PostgreSQL
Redis
Celery
Auth0
MCP
S3
Customer

"Our board wanted a credible answer to 'how are we governing AI?' that didn't depend on any one vendor's promises. 2Trust gave us that answer — with evidence attached to every message."

Dana Okafor
Chief Information Security Officer · Meridian Bank
Ready to govern

Put every AI interaction on the record.

Map one existing LLM integration onto /api/v1/ai/infer behind the Safety & Trust Layer. Stand up a governed DocumentCollection against a representative corpus. Typical pilot: 2–4 weeks.

Book demo Trust center