Uncontrolled AI use becomes a leak channel.
When every team uses a different AI tool, without approval or central oversight, personal data, contracts, patents, source code, and internal metrics can leave without any trace.
Governance and security for generative AI
Use AI models
without losing control.
SecureProxy sits between your applications and providers like OpenAI, Anthropic, Gemini, Azure OpenAI, and OpenRouter. Every call goes through policies, provider fallback, cost limits, and audit logging before reaching the model — with no app rewrite.
Real data never reaches the AI
PII, emails, and confidential data are redacted before they reach the provider. Your system gets the final response ready to use.
Automatic fallback
Primary provider down? Policy switches to the next authorized one — without touching the application.
End-to-end auditing
Call, rule, provider, cost, and latency captured in a queryable audit history.
Cost control
Per-application caps and cost per call, model, and provider.
LGPD-grade compliance
Minimization, security, and accountability you can demonstrate on every call.
Providers
- OpenAI
- Anthropic
- Gemini
- Azure OpenAI
- OpenRouter
- Mistral
- xAI
- Groq
- DeepSeek
- Ollama
- OpenAI-compatible APIs
Chapter 01
The problem.
Teams already use AI. Control over the path is what's missing.
Your teams are already using AI. The question is with what control.
LGPD and GDPR-style rules require demonstrable control.
A written policy isn't enough if the application sends data straight to the provider. Rules must be enforced in the flow and history kept for audits, legal, and incident response.
Cost and availability get scattered.
Without a central layer, every application reinvents limits, keys, fallback, and metrics. Operations loses visibility and standardization.
AI is already in operations. Control has to catch up.
The numbers point to the same problem: users adopt AI fast, often on personal accounts, and sensitive data can leak before security, legal, or IT have any visibility.
SecureProxy turns that adoption into a governed flow: central rules, authorized providers, redaction, blocking, and a history of every call.
88%
of organizations regularly use AI in at least one business function.
Source: McKinsey, 2025
75% / 78%
of knowledge workers use AI; among AI users, many bring their own tools to work.
Source: Microsoft Work Trend Index, 2024
72%
of enterprise GenAI users still operate on personal accounts at work.
Source: Netskope, 2025
26.4%
of uploads to GenAI tools contained sensitive data in a recent study.
Source: Harmonic Security, Q3 2025
57%
of exposed sensitive data was business or legal; 25% was technical data.
Source: Harmonic Security, Q3 2025
47%
of organizations analyzed already use data-loss prevention to control access to GenAI applications.
Source: Netskope, 2025
63%
of organizations analyzed had no AI governance policies in place.
Source: IBM, 2025
2%
of Brazilian revenue can be used as the base for an LGPD fine, capped at R$ 50M per infraction (parallel to GDPR's 4% / €20M).
Source: ANPD
Chapter 02
How it works.
A single control point between the application and the model.
Instead of every application deciding on its own how to handle sensitive data, keys, limits, and providers, SecureProxy centralizes those decisions on the path between the app and the model.
Queryable history. Who called, what was sent, which rule was applied, which provider answered, cost, latency, and the decision made.
- 01
Your application calls SecureProxy
Instead of calling OpenAI or Anthropic directly, the application points its AI calls to SecureProxy. The existing integration barely changes.
- 02
Content is inspected before leaving your infrastructure
Detectors find data like CPF, CNPJ, email, phone, and card numbers. Your team can also write natural-language rules — “don't send patent details” or “block customer contracts” — so the AI evaluates context before anything is sent.
- 03
The rule decides what to do
Policy can redact sensitive data before sending, block the call, flag it for review, or just record the event. It also defines providers, models, and usage limits.
- 04
The call goes to the allowed provider
If the primary provider fails, SecureProxy can try an authorized alternative such as Anthropic, Gemini, Azure OpenAI, OpenRouter, Ollama, or an internal model.
- 05
SecureProxy returns the response with a complete history
If a rule redacted data before sending, the application receives a final, ready-to-use response. SecureProxy also writes a queryable record with sent content, response received, rule applied, provider, cost, and latency.
Use AI without sending personal data to the provider.
Before calling OpenAI, Anthropic, or any other model, SecureProxy redacts PII, emails, phone numbers, and other sensitive data. The provider works on a protected version of the message. Your system gets the final response ready to use.
Your system sends
Customer João, CPF 123.456.789-00
The AI receives
Customer redacted, CPF redacted
Your system receives
Response ready for the right customer
Fault tolerance without touching the application.
Primary provider down? The policy switches to the next authorized provider and keeps the call alive.
Chapter 03
Capabilities.
What your team can configure.
Security defines which data can leave. Engineering defines which providers and models can respond. Operations watches cost, errors, and latency in one place.
Policies per application or environment
Apply different rules to production, support, internal copilots, test environments, or a specific customer — without changing application code.
Streaming-aware enforcement
Content is also checked when the response arrives incrementally, as in chats that render answers as they're generated.
Usage limits
Define how much each application can consume to prevent repeated calls, abuse, or unexpected spend on expensive models.
Cost by model and provider
See cost per call, application, model, and provider to explain spikes and split cost centers.
Protected keys
OpenAI, Anthropic, and other provider keys live outside application code — encrypted or managed in a secrets vault.
Integration without app rewrites
Your team keeps the libraries they already use. In most cases, the main change is pointing AI calls at SecureProxy.
Operations metrics
Call, cost, latency, error, and detected-data metrics are available to the operations team, including in tools like Prometheus.
Internal or external providers
Use external providers or internal models, including OpenAI-compatible APIs.
Teams and governance
Each team with its own models, budget, and policy.
A single console defines who can use which models, with what spend cap and under which set of rules. The same controls apply to applications, environments, and customers — no need to duplicate configuration per product.
Teams console
Models, budget, and policy per team
TeamAllowed modelsMonthly budgetApplied policy
Customer support
Allowed models
gpt-4o-miniclaude-haikuMonthly budget
USD 200/mo73%Applied policy
Default · mask PIIEngineering
Allowed models
gpt-4oclaude-sonnet-4Monthly budget
USD 1,200/mo41%Applied policy
Engineering · block secretsLegal & Compliance
Allowed models
azure-openai (BR)Monthly budget
USD 150/mo88%Applied policy
Critical · block regulated dataResearch & Data
Allowed models
all providersMonthly budget
USD 500/mo12%Applied policy
Research · flag and audit
Each team only sees what it can use. Billing and blocking happen at the edge — not in the app.
A different policy for each risk.
Each policy in the console bundles a concrete set of rules. Four examples of what different teams actually run day-to-day, without duplicating logic across products.
Customer support
Chatbot with customer data
- Applied rule
- Mask PII, phone, and email; block card data; record calls that contain personal data.
- Provider, cost, and operations
- Fast models with per-conversation cost limits.
Internal copilot
Search across contracts and documents
- Applied rule
- Block confidential clauses, customer contracts, and information flagged as restricted.
- Provider, cost, and operations
- Approved external provider or internal model for restricted documents.
Engineering
Code and incident assistance
- Applied rule
- Prevent leakage of keys, secrets, proprietary code, and details of undisclosed incidents.
- Provider, cost, and operations
- Models allowed per project, with fallback when the primary fails.
Finance and Legal
Reports, M&A, and contract analysis
- Applied rule
- Natural-language rules to block projections, negotiation details, patent info, or regulated data.
- Provider, cost, and operations
- Full audit log for traceability, cost, and decision review.
Chapter 04
Trust.
Isolation, auditing, and demonstrable compliance.
Isolation and key management don't depend on you.
SecureProxy enforces separation between customers, areas, and environments by default. Keys, administration, and AI traffic live on different planes, ready for stricter network rules.
Isolation across customers, areas, or environments
Each organization only sees its own applications, policies, and records. Isolation is enforced down to the database, reducing the risk that one customer or internal area accesses another's data.
Keys outside application code
Provider keys are centralized in SecureProxy, encrypted at rest or managed by a secrets vault like Vault/OpenBao. If a key needs to be rotated, the application doesn't have to be redeployed.
Administration separated from AI traffic
AI calls, administration, and metrics live separately. That makes stricter network rules easier and reduces what has to be publicly exposed.
Every call becomes queryable evidence.
Sent content, applied rule, responding provider, cost, and latency are all captured in an auditable history and available on the dashboard for audit, investigation, and accountability — without depending on logs scattered across applications.
Audit dashboard
What the admin sees.
The pipeline writes every call into the audit log. The dashboard turns that into queries: KPIs over time, usage trends, and the list of recent calls — team, model, applied policy, status, and cost on a single line.
Audit dashboard
Every call with its team, model, policy, and cost
Calls
24,871
12%vs last month
Total cost
USD 487.32
8%vs last month
Blocks
142
18%vs last month
Avg latency
412 ms
4%vs last month
Calls per day
30d
TimeTeamModelPolicyStatusCost (USD)
Time
14:32:08Team
SupportModel
gpt-4o-miniPolicy
DefaultStatus
SENTCost (USD)
0.004Time
14:31:47Team
LegalModel
azure-openaiPolicy
CriticalStatus
MASKEDCost (USD)
0.012Time
14:31:22Team
EngineeringModel
claude-sonnet-4Policy
EngineeringStatus
BLOCKEDCost (USD)
0.000Time
14:30:55Team
ResearchModel
gemini-1.5-proPolicy
ResearchStatus
SENTCost (USD)
0.008Time
14:30:31Team
SupportModel
claude-haikuPolicy
DefaultStatus
MASKEDCost (USD)
0.002Time
14:30:04Team
EngineeringModel
gpt-4oPolicy
EngineeringStatus
SENTCost (USD)
0.015
Filter by team, policy, status, or time window. Each row points back to the original call in the audit log.
Compliance isn't just blocking PII. It's demonstrating control.
LGPD (and GDPR-style regimes) require practices aligned with purpose, necessity, security, prevention, and accountability. SecureProxy helps apply those controls at the exact point your application talks to AI.
Read the official LGPD principles (Brazil)Minimization and necessity
Send to the provider only what policy permits. Personal data, contracts, source code, patents, and other secrets can be removed, masked, or blocked before they leave.
Security and prevention
Centralize keys, allowed providers, usage limits, and content rules. Teams keep using AI — with controls applied on the path.
Accountability
Maintain a queryable history of sent content, response received, rule applied, provider, cost, and latency to support audit and investigation.
Chapter 05
Where it runs.
Managed by us or inside your perimeter.
Pick the model that matches your risk: a dedicated environment we operate, or an install inside your perimeter when data can't leave.
Managed isolated
We run it. The environment is yours alone.
- ✓Dedicated environment for your organization
- ✓No execution resources shared with other customers
- ✓Assisted setup of providers, applications, and policies
- ✓Data processing agreement and security documentation
- ✓Managed updates, backups, and operation
On your infrastructure
Inside the network your team controls.
- ✓Deploy on Docker/Compose or your own servers
- ✓Support for restricted environments or no direct internet egress
- ✓Call history and admin dashboard stay inside your perimeter
- ✓Traefik and automatic TLS included in the local install package
- ✓Integrates with secrets vaults like HashiCorp Vault or OpenBao
FAQ
Questions we hear a lot.
The best demo uses a real case: which application calls AI, what data can't leave, what security or LGPD obligations need to be met, and which providers are allowed to respond.
OpenRouter helps you reach many models. LiteLLM helps you normalize technical calls. SecureProxy adds governance: before the call reaches the model, security and compliance rules are applied; when sending, the allowed provider is used; after, a queryable history of what happened is kept.
No tool guarantees compliance on its own. SecureProxy helps apply technical controls that matter for an LGPD (or GDPR-style) program: minimizing what's sent, blocking or masking sensitive data, per-organization separation, key management, and a queryable audit history.
It depends on the provider you choose. If the call goes to OpenAI, Anthropic, or another external provider, SecureProxy removes, masks, or blocks sensitive information beforehand. When nothing can leave the network, you can point at internal providers like Ollama or any OpenAI-compatible API.
It depends on call size and which rules are active. A simple rule, like masking a PII number, is different from a contextual rule that uses AI to evaluate content. In pilots we measure with real traffic and show the impact before going to production.
OpenAI, Anthropic, Gemini, Azure OpenAI, OpenRouter, Mistral, xAI, Groq, DeepSeek, and Ollama. You can also connect internal or regional providers that expose an OpenAI-compatible API.
No. Structured detectors cover common personal data. For context-dependent information, you can write rules like “don't leak patent info,” “don't expose negotiated prices,” or “don't reveal internal financial data.” The rule can block the call, redact passages, flag for review, or just record.
Policy can define alternates. If the primary fails, the call can go to another allowed provider like Anthropic, Gemini, Azure OpenAI, OpenRouter, or an internal model. The application keeps calling the same SecureProxy endpoint.
On managed isolated, we provision the environment and set up providers, applications, and initial policies with your team. For on-prem installs, we ship the deployment package with Docker/Compose, Traefik, and secrets-vault integration when needed. Pilots begin with a real AI call flow.
Technical demo
Bring a real call. Watch governance happen.
We'll show the application calling SecureProxy, the rule redacting or blocking sensitive data, routing picking the provider, and the history capturing the decision for security, legal, and operations.