State-of-the-art AI. Your data encrypted, even during processing.Compatible with OpenCode, Claude Code, Codex CLI, Codex Desktop, and our API.
PRIVACY, BUILT IN
Your data stays encrypted on its way to the model and back, and protected while the model works. You don’t have to trust a privacy policy.
CONVENTIONAL AI
The lab can read and store both sides of your conversation. Copies can spread to backups and vendors, with access for employees and hosting companies.
Even a zero-retention policy asks you to trust that whole chain. Policies can be broken, by mistake or by choice.
WITH SERENCE
An unreadable stream connects you to the protected model. Serence, hosting companies, vendors, and employees cannot read your prompts or answers, or keep readable copies.
No readable copies. No content to train on. Cryptographically enforced.
PRIVACY IS FOR EVERYONE
Give your coding assistant the context it needs, without giving a lab your codebase. Work with proprietary algorithms, internal systems, and sensitive data inside private inference.
function nextBreakthrough(yourIdeas) {
return /* yours, and yours alone */
}Review agreements, compare clauses, and reason through sensitive matters. Protect the contents of client documents during inference, including information you would never send to a public chatbot.
Your client’s confidence.
Kept in confidence.
Bring email, calendar, and Drive context into private AI workflows. Plan, synthesize, and make decisions without letting an inference provider read your communications or company strategy.
Your next move.
On a need-to-know basis.
Connector permissions and data handling must preserve the protected path. Discuss your integration requirements with us.
Connect your workflow privatelyUse cyber-capable models to pentest and red-team infrastructure you own or are authorized to test. Take security into your own hands without waiting for admission to a lab’s restricted security program.
Your infrastructure.
Your security program.
Discuss models suited to authorized security work without blanket refusals. Model behavior and access depend on the deployment.
Discuss your security workflowExplore hypotheses, analyze results, and develop new proofs without handing over unpublished work. Your research should advance your team, not become training material for someone else’s model.
The next breakthrough
starts in private.
THE RISK ISN’T HYPOTHETICAL
OpenAI disclosed a bug that let some ChatGPT users see other users’ conversation titles and possibly the first message of a new conversation. Limited billing details were also potentially exposed for some active Plus users.
Wiz found an exposed DeepSeek database with more than one million log entries, including plaintext chat history and API secrets. DeepSeek secured it after disclosure. Exposure was verified; malicious theft was not established.
Hacktron compromised OpenAI employee accounts. Those accounts opened access to its codebase, email, and other internal services. Even frontier labs get hacked. Your privacy shouldn’t depend on their accounts staying secure.
OpenAI initially said it was unlikely, but could not rule out usage-derived data helping improve its models. It later reported ruling out influence from Buckmaster’s Codex prompts in the preceding two months, including through training.
This is not evidence of copied research. It raises a bigger question: why should unpublished work enter a provider’s data pipeline at all?
Anthropic reported that some proxy services recorded Claude conversations and sold them to AI labs. Its report describes SenseTime purchasing transcripts harvested through third-party apps and routing services.
Inference is a model turning your request into an answer. Serence keeps that work inside a protected environment, with your content encrypted on the way in and out. We run the service without being able to read your content. How it works ↗
Those policies promise what a provider will do with data it can still read. Serence protects your prompts and answers with cryptography: we cannot read them, store a readable copy, or use them for training.
A client on your computer verifies the protected environment and approved software using cryptographic proof before sending your encrypted data. If verification fails, nothing is sent. How it works ↗
No. The hosting provider has no read access to the data you send us or the answers you receive.
We and our intermediaries cannot read anything you send to Serence for processing, including emails, calendar events, files, and prompts—or the answers you receive. Google and other connected services remain responsible for securing their own services and the data they hold.
Yes. Use OpenCode, Claude Code, Codex CLI, or Codex Desktop—or use our API directly in your own software.