SaaS & Tech
Ship AI Features Without Sending Your Customers' Data to the Model
- 01
Request from your application
- 02
Sensitive values masked
- 03
Forwarded to the AI provider
574
Built-in detectors for personal data and secrets
47
Country packs, checksum-validated where the format has one
9
AI providers behind one gateway, plus the desktop agent and proxy
Where customer data meets AI in a software company
Your support team uses AI to draft replies to tickets, and every ticket has a customer's name, email, and sometimes a screenshot of their account. Your product team ships an assistant that answers questions over the customer's own documents. Your engineers run Claude Code or Cursor against a repo with a `.env` file in it.
Then an enterprise prospect sends a security questionnaire asking which sub-processors receive their data and whether it is used for AI. Each AI provider you call with customer data is a sub-processor you have to list, and every new one can trigger a notice under your DPAs.
Maskflare gives you one place to control that: the gateway and SDKs for product traffic, MCP masking for agents, hooks for coding agents, and the desktop agent for staff on AI chat.
Where saas & tech is actually exposed
5 failure modes specific to this environment, not a restatement of general security advice.
01
Customer data in support tickets
An AI-drafted reply sends the whole ticket to the provider. Through the gateway, names, emails, and order numbers leave as tokens and come back as real values in the draft the agent reviews.
02
Logs and stack traces
Engineers paste logs into AI to debug. Logs carry user emails, IP addresses, session tokens, and API keys. The credentials pack catches 232 secret formats and a block rule refuses the worst ones.
03
RAG over customer documents
Indexing customer files puts their personal data into embeddings and every retrieved chunk into a prompt. Deterministic tokens keep retrieval working, and embeddings requests fail closed if masking can't complete.
04
Coding agents with shell and file access
An agent that reads `.env`, runs `cat` on a config, or calls a database tool can send secrets and customer rows to the model. Hooks for Claude Code, Cursor, and Codex CLI mask or block them in prompts and tool calls.
05
Agents calling MCP tools
An agent that queries your CRM through an MCP server pulls customer records into its context. The `/v1/mcp/{server}` route masks tool arguments and results before the model sees them.
Who this page is for
- CTO / VP Engineering
- Wants AI features in the product without a privacy review stalling each one.
- Head of Security
- Answers security questionnaires and needs AI data flows to be one of the easy questions.
- Privacy and legal
- Maintains the DPA and sub-processor list and wants fewer customer-data flows to disclose.
- Platform engineering
- Wants one integration in front of nine providers, not a masking library in every service.
- Support operations
- Wants AI-drafted replies that address the customer by name without sending the ticket to a model in clear text.
How a team would actually run this
Putting product AI traffic behind the gateway
- 01 Change the SDK base URL to `https://<your-gateway>/v1/openai` or `/v1/anthropic`. No other code change.
- 02 Turn on the US PII (CCPA/CPRA) or GDPR profile, plus Exact Data Match for your own customer and account IDs.
- 03 Switch on response restoration for the gateway key, so streamed answers show real values as they arrive.
- 04 Use `mf_dev_`, `mf_stg_`, and `mf_live_` keys so rule changes are tested before they reach production.
Keeping secrets out of coding agents
- 01 Install `maskflare-hook` from the Python SDK and register it with Claude Code, Cursor, or Codex CLI.
- 02 In Claude Code, prompts with secrets are blocked and Bash, WebFetch, Write, Edit, and MCP tool input and output are masked. Cursor and Codex CLI hooks cover less, because their hook APIs expose less.
- 03 Set private keys and cloud credentials to block; set lower-risk tokens to mask so the agent keeps working.
Answering the security questionnaire
- 01 List which flows send customer data to which provider, and which of them run through Maskflare.
- 02 Export the compliance coverage report and audit log as evidence of the rules in force.
- 03 List Maskflare itself as a sub-processor, and state its current certification status from the Trust Center.
Which modules apply, and why
Each links to the module page, where the boundaries and development status are set out in full. Or start at the Maskflare platform overview.
For your apps
Maskflare AI Gateway
A base-URL change puts product AI calls to nine providers behind one set of rules, with streaming, embeddings, and Batch uploads masked.
For your apps
Maskflare SDKs & API
Python and TypeScript SDKs mask and unmask text inside your own pipelines, with a LiteLLM guardrail for teams already on LiteLLM.
For your agents
Maskflare Coding Agents
Hooks for Claude Code, Cursor, and Codex CLI keep secrets and customer data out of agent prompts and tool calls.
For your agents
Maskflare Agent & MCP Security
Masks MCP tool-call arguments and results before the server or the model sees them, and returns a JSON-RPC error on a block rule.
Detection engine
Maskflare Exact Data Match
Masks your own customer IDs, account numbers, and tenant names from keyed hashes, with no clear values sent.
For your people
Maskflare Workforce
The desktop agent covers support and sales staff on AI chat sites, rolled out through Intune, Jamf, or Group Policy.
Regulatory context
Readiness and relevance, not certification. Nothing here is a claim to hold an audit or authorization we do not have.
- GDPR Art. 28
- You are usually a processor for your customers. AI providers you call with their data are sub-processors to list. Masking changes what you disclose: tokens rather than customer data, though Maskflare itself is still a sub-processor.
- SOC 2 (yours)
- Your auditors and customers ask how confidential data is protected in AI flows. Maskflare's rules, audit export, and maker-checker approvals give you evidence. Maskflare's own SOC 2 Type II and ISO 27001 audits are in progress.
- CCPA / CPRA
- A one-click US PII profile covers the personal information categories it detects in text. Contract terms with providers still matter.
- EU AI Act
- Obligations depend on what your AI feature does. Maskflare reduces personal data in prompts; it doesn't classify or assess AI systems.
What Maskflare does not do
- Maskflare's SOC 2 Type II and ISO 27001 audits are in progress; ask for the security documentation when you list it as a sub-processor.
- Coding-agent hooks cover what each agent's hook API exposes; check the capability matrix for your tools.
- Realtime voice: text events are masked, audio is not.
- Free-text personal details that match no pattern, list, dataset or named entity can be missed, so test with your own documents.
4 questions to ask any vendor here
Including us. If our answer is worse than someone else's, you should know that before you buy, not after.
- 01Does it handle streaming, embeddings, and tool calls, or only plain chat requests?
- 02What does it add to latency on a typical request? (Our figure: about 2 ms median on a 10 KB chat request with default rules.)
- 03Does it store prompts or responses, and can it run in our own environment?
- 04What happens when masking fails: does the request go out unmasked or fail closed?
Workflow guides for this industry
How a team applies each of these, including the contrast with the tool it replaces.
Definitions worth agreeing on first
SaaS & Tech questions we get asked
Does using Maskflare remove AI providers from our sub-processor list?
Not automatically. If only tokens reach the provider, you have a stronger answer about what it receives, but whether it still counts as a sub-processor is a legal call for you. Maskflare itself is a sub-processor. Not legal advice.
How much latency does it add?
About 2 ms median on a 10 KB chat request with default rules (4.3 ms p99), and around 10 ms with all 574 detectors on.
Does RAG still work on masked data?
Yes. Embeddings use deterministic tokens, so the same value always gets the same token and queries still match the index. If masking can't complete, the embeddings request fails closed.
Which coding agents are supported?
Claude Code, Cursor, and Codex CLI through `maskflare-hook`. Claude Code gets the widest coverage, including Bash, WebFetch, Write, Edit, and MCP tool calls.
Can it run in our own cloud?
Maskflare is EU-hosted by default, and on-premises deployment is available for enterprise. Prompts and responses are never stored either way.
See it on your own data
Start a free trial or talk to us with the identifiers and AI workflows your saas & tech team actually handles. Bring the case that makes it hard, because that is the part worth testing.
See it on your own data
Try Maskflare free.
Bring your hardest prompts.
28-day free trial or 10,000 requests, whichever comes first. Point one app at the gateway or enroll a few laptops,
and watch what gets masked. Or talk to us first.
- SOC 2 Type II: In progress
- ISO 27001: In progress
- EU hosting by default · on-premises available
- Prompts and files never stored
- Never used to train AI models
- GDPR · nDSG · DPA available
- Trust center →