CONFSEC gives your AI the following guarantees
Prompts are never logged or retained
Prompts are never used for AI training
Prompts are never sent to third parties
Prompts are completely anonymous
Prompts are verified to not be tampered with
System operators do not have access to private computation
With 8 key technologies to achieve these guarantees
IETF standard that offers a mechanism to make HTTP requests to a server without revealing to the server the true requester
Digitally sign something without seeing the content and without linking the later revealed content to the earlier signature
Mechanism for a client (“relying party”) to ask a server (“attester”) to prove its identity and what it is running
Secure system that keeps sensitive data hidden and inaccessible from the main operating system
Tamper-proof append-only public record of software packages available for review
Process to ensure booted and running software is trusted + a stripped down version of linux that is more difficult to hack or modify
Virtual machine initialized inside a hardware-based trusted execution environment (TEE) so that the host cannot see inside the virtual machine
Verify that the running binary is the same as what the compiler generates from the source code itself
Together, these technologies create CONFSEC, a private AI inference wrapper
CONFSEC meets the following requirements
User data is used exclusively to fulfill the user’s request, not available to anyone, and must not be retained, including via logging or for debugging.
Code and infrastructure -- not convention or policy -- is designed to make certain things impossible and to ensure the system behaves correctly
An attacker cannot compromise data that belongs to specific users without attempting a broad compromise.
Our staff cannot bypass privacy guarantees for any reason.
Security researchers must be able to verify the security and privacy guarantees and must be able to verify that the software running in production is the same as the software they inspected.