You lock your front door without having anything to hide. Every time you use AI, a copy of something you know travels to a computer you do not own. This walkthrough shows who can read that copy, how AI agents raise the stakes, and how to use the newest models without leaving the door open.
> nothing you select here is sent or saved. it all stays in this browser tab.
People paste remarkable things into AI tools because it saves time. Select the ones you, your family or your staff might paste, and see what each one hands over.
// Select any that apply. Nothing is recorded.
The question is never “do I have something to hide?” It is “who ends up holding a copy?” The next chapter answers that.
We use Venice AI’s four privacy modes as the map, because Venice spells out the trade-offs in plain terms. Select each rung to see who can read your prompt, what backs the promise, and what it costs you.
Every rung answers the same three questions.
The service that passes your prompt along. Can it read it?
The computer that runs the model. Can its owner read or keep it?
What backs the promise: nothing, a contract, hardware or mathematics?
Hides who you are, not what you said.
Passes your prompt along and does not store it.
An outside frontier model provider. It reads your prompt, and you should assume it keeps it.
Your identity is hidden from the provider. Nothing protects the words themselves.
You get the leading frontier models: GPT, Claude, Gemini.
You give up control of the content. The provider sees and likely saves it.
Nothing kept, on a promise.
Passes your prompt along and does not store it.
Venice’s own GPUs, or partners signed up to zero data retention. It processes your prompt and does not keep it.
A contract. You are trusting Venice and its partners to honour it.
You get no retention and a wide choice of models, including Grok.
You give up the ability to check. It is trust, not proof.
Sealed hardware, with a receipt.
Still receives your prompt as readable text, then passes it into the enclave.
A hardware-isolated enclave run by outside partners. The operator of the machine cannot see your prompt.
Remote attestation: a cryptographic report that the model really is running inside sealed hardware. It can be checked.
You get privacy backed by hardware instead of a promise.
You give up speed and choice: responses may be slower and fewer models are available. Text models only.
Locked before it leaves you.
Sees only scrambled text. It cannot read your prompt.
The sealed enclave is the only place your prompt is ever unlocked.
Encryption on your own device, plus the enclave’s attestation. Venice describes this as private from all parties, itself included.
You get the strongest protection on the ladder.
You give up some speed and most of the model list. Text models only. Venice switches off its own web search and memory here, but an agent such as Agent Zero runs both itself, so you keep them.
// As described by Venice AI at venice.ai/privacy and docs.venice.ai, read 10 Nov 2026. Modes and models change; check the source.
Notice the pattern: every step up buys proof and costs capability. The top rung cannot search the web. The bottom rung has the newest models. Using AI well means choosing the rung for the job, not living on one.
Too low and you are exposed. Too high and you gave up the best models for nothing. Choose where each job belongs. “Encrypted” means rungs 3 and 4.
Summarise this morning’s public market news.
Read a tax return and explain it line by line.
Compare the published fees of two index funds.
Find the duplicate charges in a card statement.
Talk through your goals and rough net worth, with no name attached.
Draft a letter to your attorney about your trust.
// These are our rules of thumb, not a standard. The principle is the point: the more a document could help someone impersonate or target you, the higher it goes.
The newest AI tools do not just talk. They read your inbox, open web pages, fill in forms and make payments. That is the real advance, and it brings a risk no privacy setting covers: an agent can take instructions from whatever it reads.
Hello,
Your November service invoice is attached. Amount due: $480 by Nov 30.
Thank you,
Harbor Pool Service
// Illustrative example. Not a real company or message.
You see an ordinary invoice. Ask your AI agent to “handle my invoices” and it reads every character.
// Source: OWASP Top 10 for LLM Applications 2025, LLM01 Prompt Injection.
It starts with everything switched on. Switch off what should wait for your yes.
// OWASP calls the underlying fault excessive agency: too many functions, too many permissions, too much autonomy. Its fixes are the ones on this panel: give the agent the minimum, and require a person’s approval for anything high-impact. Source: OWASP LLM06:2025.
Encryption protects what you send. It does not check what an agent reads, or what it does next. Staying safe takes both: the right rung, and an agent on a short lead.
The same technology that writes your reports can imitate a person. In December 2024 the FBI warned that criminals generate short audio clips of a loved one’s voice to impersonate a close relative in a crisis and ask for money immediately.
An unfamiliar number. Your daughter’s voice, crying. There has been an accident, and she needs $15,000 wired within the hour.
// Source: FBI Internet Crime Complaint Center, alert I-120324-PSA, 3 Dec 2024. The scenario is illustrative.
AI is neither a miracle nor a menace. The same capability shows up on both sides of the ledger, and it grows on both sides every month.
// Right-hand column: FBI IC3 alert I-120324-PSA, 3 Dec 2024.
Anyone who tells you one setting solves this is selling something. The honest answer is layers: the right rung for each job, narrow permissions for agents, and a person who checks the result.
Client files are end-to-end encrypted by default. When a question would benefit from a frontier model, you can choose to add one. Names and account numbers are stripped first, and the choice is yours each time.
Default: the strongest protection on the ladder. The cost is a shorter list of models and some speed. Web search and memory still work, because Agent Zero runs them itself rather than relying on Venice.
Enhanced: the newest frontier models work on your question. The cost is that a provider sees the stripped text and may keep it.
// Everything shown on this site is a case study built from sample documents. No client’s files appear here.
None of these needs a technician. All of them work whatever AI tools you use.
Your shield score measures what you explored on this page, not how secure you are. That takes a real review. A 45-minute introductory session is $100.