Launching the privacy focused AI chat platform

Generative AI is changing how we all create content and being integrated everywhere, but as the companies producing these models reach the limits of what is possible with open data, they increasingly rely on the information you are giving them to train their future models. This presents a bit of a privacy nightmare for individuals and companies who want to use AI but value their privacy and confidentiality.

Employees of companies such as OpenAI and Google can - and do - access your messages you send through their products such as ChatGPT and Gemini.

On top of that, your chat histories could be accidentally leaked (this has already happened to OpenAI), stolen by hackers, sold by the software you're using or acquired by a bigger company (like when Google acquired Nest).

Are you going to continue risking your personal data?

If you have the technical knowledge and a few hundred/thousand dollars to spend on hardware, it is possible to find open source software that you can setup and run yourself, ensuring your data never leaves your own private garden. For the truly paranoid, running everything yourself is the best option.

But what about people who:

1. Want to use generative AI;
2. But also want to preserve as much of their privacy as possible;
3. Can't, or don't want to, run and maintain AI models or software themselves;

Cognos is the privacy-focused alternative that tries to offer the convenience of cloud AI but encrypts your chats so that only you can access them.

Demo of our Beta

It's nice to know what is being offered without having to sign up so please find some screenshots below so you know what to expect. This is a beta project though so there is only basic chat features at the moment.

Feature list

The basic product is now live with the following features (but there is lots more planned).

Security

Cognos can't claim to secure your messages without explaining the security involved and demonstrating that there is zero unauthorized access to your messages.

The simple version

It's important to note that this is not end-to-end encrypted (such as WhatsApp or Signal) as the AI models you choose to interact with - currently - need to receive your messages in plain-text.

Instead, a layer of encryption is added on top like ProtonMail/Tuta/Skiff do for email. If you're familiar with such services, they take emails to you which are unencrypted and encrypt them so only you can read them. When you send an email to a non ProtonMail account, this is also sent unencrypted.

Cognos does something similar, meaning that the messages you send to our servers and to the AI are in plain text (although the connection between you, us and the AI provider is secured with HTTPS to prevent eavesdropping) but after that they are encrypted so you are the only person who can decide who can access your messages.

The technical version

You start by creating a user account on our external identity provider (IDP) Ory who are ISO 27001 and SOC 2 Type II certified.

When you first create an account you are asked for a "Vault Password". This never leaves your device and is used to symmetrically encrypt and decrypt your secret key. To make sure this has enough entropy we first hash your vault password with Argon2id and the parameters recommended by OWASP. This turns your Vault Password into an alphanumeric, fixed length string that will be used as a NaCl Secretbox key. To combat rainbow table attacks your user UUID (generated by Ory) is used as the Salt.

After hashing your Vault Password a Curve25519 public/secret key pair is generated on your device. This needs to be stored on the server so it can be shared between different devices so your hashed Vault Password is used to encrypt the secret key using an NaCl Secretbox - XSalsa20 cipher combined with Poly1305 message authentication. Your plaintext public key and encrypted secret key are now sent to and stored on the server for future use.

When you create a conversation, another unique Curve25519 public/secret key pair is generated for that conversation. The conversation public key is also sent to the server in plaintext however this time a shared key is generated from the conversation public key and your secret key. This shared key is used with an NaCl Box to encrypt the conversation secret key with public-key cryptography. While not supported at this time, in the future this will allow you to invite other users to a shared, encrypted conversation. Each user will have access to the conversation secret key which is encrypted with their personal secret key.

Sending a message involves transmitting the message to the server in plaintext over TLS encrypted connections. Your plaintext message is encrypted with the conversation public key and saved before being forward to your AI model of choice, again in plaintext over an TLS encrypted connection. When the AI model has generated a response, this is returned to our server. This response is also encrypted with the conversation public key and saved before being forwarded back to you.

No plain text copies of your messages or the AI responses are stored or logged.

Using this approach, when you next sign in to access your messages we:

1. Hash your Vault Password with Argon2id salted with your user UUID;
2. Retrieve your Curve25519 public and encrypted secret key pair;
3. Decrypt your secret key using your hashed Vault Password;
4. Retrieve the Curve25519 public and encrypted secret key pair for the conversation you are trying to access;
5. Use your secret key and the conversation public key to decrypt the conversation secret key;
6. Retrieve the XSalsa20Poly1305 encrypted messages for the conversation;
7. Use the decrypted conversation secret key to decrypt all messages;

Your messages are only readable by you and - in the future - the people you choose to share your conversation with. Cognos can't read your messages or use them to train any AI. In the event of a data leak nobody can decrypt your messages. Hackers cannot decrypt your messages (although there is a risk to you if you are using an insecure Vault Password).

AI Model list

You can currently choose from the following providers and models:

• OpenAI:
  • GPT-3.5-Turbo
• Cloudflare:
  • Meta/LLama3-8B-instruct
  • Mistral/Mistral-7B-instruct
  • Deepseek-AI/Deepseek-math-7B-instruct
  • Qwen/Qwen1.5-7B-Chat

We are working to add and release these during the Beta:

• Google:
  • Gemini-1.5-Flash
• Anthropic:
  • Claude3-Haiku

When going live there will also be additional larger models and others will be continuously added to our offering, based on user demand and new releases from providers.

Privacy policies of AI providers

You may have concerns about sending your data to our upstream AI providers and that is understandable. In the mid-to-long term we want to have our own hardware to run models on Cognos owned infrastructure but this is expensive and not possible to do this from the beginning.

For those who don't know, using the API - like Cognos does - of services like OpenAI often gives you immediate privacy benefits versus using their products like ChatGPT, as they will not use it for their training data. Unfortunately however:

[...] data submitted to OpenAI’s services through automated content classifiers and safety tools, including to better understand how our services are used. The classifications created are metadata about the business data but do not contain any of the business data itself. Business data is only subject to human review as described below on a service-by-service basis.

You have the choice about where to draw the line when it comes to your data. Perhaps you are happy to have some messages sent to OpenAI or Google. Perhaps you prefer to er on the side of caution and use only open source models hosted on Cloudflare, Lightning AI or similar. It's up to you and we'll give you as many options as possible.

A little about me and why I'm doing this

I'm Ewan a Scottish software engineer living in Luzern, Switzerland and I'm currently working full-time building projects to have an impact.

Cognos was built to satisfy my own itch of wanting to use generative AI in a more privacy friendly way. While I would have the skills and knowledge to build my own AI setup I have decided not to spend the time or the money. It's also not a valid option for my non-technical friends or family.

This is my attempt to bring confidentiality to AI chats and it is currently being used by me and my family. I have no interest in what any of my users are talking to AI about.

If you want to talk to me about Cognos I'm happy to hear from you via:

• Email: ewan@cognos.io
• Threema: NM4AVD9N

Frequently asked questions (FAQ)