How Saibre works

Corporate networks have inspected TLS traffic for two decades. The technique is identical: a certificate authority the machine trusts, sitting between the application and the internet, so encrypted traffic can be examined before it leaves.

Saibre changes who holds the keys. The CA is generated on your device at install time — not issued by us, not shared between users, and never transmitted anywhere. The private key is encrypted with your operating system's own keystore: DPAPI on Windows, Keychain on macOS.

The certificate is capped at roughly 27 months of validity. Corporate inspection CAs are routinely issued for ten years. A shorter life means a smaller blast radius if a machine is ever compromised.

Superfish shipped on Lenovo laptops in 2015 with the same CA private key on every machine — anyone who extracted it once could intercept traffic for every affected user. Kaspersky intercepted TLS for the whole system and weakened certificate validation while doing it. Both became public scandals, and they should have.

Saibre's design differs on each point that made those failures. The CA is generated per device at install time; there is no key to share because no two installs have the same one. Interception covers only a published AI-domain allowlist, not the whole system. And nothing is silent: you run the installer yourself, and every part of what it sets up is inspectable afterwards.

If you're evaluating us critically, this is the right comparison to make. We'd rather make it for you than have you wonder whether we've thought about it.

Saibre routes a published allowlist — 65+ domains across 20+ AI apps — through inspection. Your banking, your email, your browsing, your work VPN: none of it is inspected, and none of it is decrypted. It goes direct.

This is not a promise in a policy document; it is enforced by a proxy auto-config file you can read yourself at 127.0.0.1:3129/proxy.pac. If a domain isn't in that file, Saibre never sees the traffic.

Redaction and attack detection run entirely on the machine. There is no Saibre server that receives your prompts, your files, or your secrets — because there is no server in that path at all.

This is the specific design decision that separates Saibre from cloud AI-security products, which inspect your content on their servers. That architecture requires you to send them everything you were trying to protect. We reject it.

Saibre reports its own coverage per app: protected for all users of the machine, partial, or none. When a path can't be inspected — a pinned app, a browser profile we haven't configured — you see that, instead of assuming protection you don't have.

No security product likes admitting where it's blind. We'd rather you learn it from our interface than from an incident.

If inspection fails for any reason — a malformed payload, a parser it cannot safely handle — the traffic is blocked, not passed through. A security tool that leaks when it breaks is not a security tool.

And if the Saibre agent itself goes down, a watchdog blocks direct AI traffic until protection is back up, on both Windows and macOS. You will notice that Saibre stopped working, because your AI tools stop working — rather than quietly losing protection without knowing it.

Policy is cryptographically signed. A configuration file modified on disk — by malware, by a curious process, by anyone — fails verification and resets to safe defaults rather than being honoured.

Every block tells you exactly what was caught and which detectors fired, with a one-click bypass when you disagree. On managed devices, that bypass is refused by the agent's own policy engine, not by a greyed-out button in the interface.

Every bypass is written to a local audit log that you own and can read.