Correcting AI Agent Security Distortions

Aspheric.ai

> Don't trust the AI. Make trust irrelevant._

govshell — zsh
dirty-shell [CONTAMINATED INPUT] scrub-pipe [ASPHERIC CORRECTION] clean-shell [STERILE OUTPUT]

Bring AI Into Alignment.

Your AI needs passwords to be useful.

It also can't keep a secret.

AI agents write code, send emails, and manage databases. To do any of this, they need access to real credentials and private data. The industry's answer is "trust the AI more." Ours is different: we built a clean room. Before any real data reaches the AI, we swap it for format-authentic fakes. The AI does its work. When results come back, we swap the real values in.

"The AI can't leak what it never had."

Two shells. One pipe.

Your shell is dirty. The AI's shell is clean. scrub-pipe sits between them.

dirty-shell

The real environment

Your terminal, your credentials, your data — all real. This is the contaminated side. Real passwords, real keys, real private data live here.

scrub-pipe

The transform layer

Data flows through and gets scrubbed. Real values go in, format-authentic fakes come out. It watches, logs, and transforms everything between the two shells.

clean-shell

The AI's clean room

Everything looks real. Nothing is. Every piece of sensitive data has been replaced with a format-authentic fake. The AI works here and never knows the difference.

GovShell. One command.

The governor for AI agents. Open-source, terminal-native, zero-config. Inspect every line. Contribute improvements. Deploy with confidence.

quickstart
# Launch the clean room
$ govshell bash
> Ready. All outbound AI traffic is now governed.
View on GitHub