Bastian · AI safety

AI agents now hold your company’s keys. And they obey what they read.

Bastian stands between your staff’s AI agents and everything they can reach: mail, files, accounts. It keeps your passwords away from the agents, sets traps that only a machine would touch, and watches your own logs. Built by AVA Digital for companies of 10 to 300 people.

33 of 39
AI agent runs caught
0 of 140
Staff sessions flagged by mistake
0 of 12
Agents reached the data with blocking on
Sealed lab · made-up data · five test rounds, October 2026
In one paragraph

Bastian is AVA Digital’s AI safety layer for companies whose staff use AI agents. Agents never see your passwords, anything final waits for a person’s tap, decoys expose any machine that goes looking, and your own access logs are checked for machine behaviour. Start with an audit, from €2,500.

The problem

One hidden line is enough.

01

An email arrives with a line no person can see.

02

Your AI assistant reads it and follows the order.

03

Client files leave the company. Nobody clicked a thing.

An illustration of how it happens, not a reported case. In September and October 2026, Meta, OpenAI and Anthropic each launched agents that act for you: they read your mail, open your files and use your accounts.

What Bastian does

It stands between your agents and everything they can reach.

Holds the keys

An agent gets a short-lived pass for one task, never the password. The password stays in a vault. Anything that cannot be undone, such as sending, paying or deleting, waits for your tap on your phone.

Sets the traps

Fake logins, keys, files and hidden notes that no person would ever use. Anything that touches one is a machine, and you know at once.

Keeps the watch

Your own access logs, scored session by session and account by account, for the habits a machine has and a person does not.

Starts with an audit

We map where your agents hold the keys, test your exact setup in our sealed lab, fix what we find, then watch it every month.

Tested before it reaches you

Proven on a made-up company first.

33 of 39
AI agent runs caught
0 of 140
Staff sessions flagged by mistake
0 of 12
Agents reached the data with blocking on

Five test rounds in our sealed lab, October 2026, on a made-up company with made-up data. No real data, no real targets.

What it does not catch alone: a slow, quiet agent that avoids every trap. Scoring the same account across sessions caught that agent in round 5. We tell you the limits before you buy.

Prices

Fixed prices. No surprises.

Audit · one-off

€2,500
Up to about 50 staff
Buy the audit
€4,000
Up to about 300 staff
Buy the audit

Traps and Watch · monthly

€490 a month
Up to about 50 staff, one site
Subscribe
€990 a month
Up to about 300 staff, several sites
Subscribe

Gate add-on · monthly

€190 a month
Up to 10 agents
Add the Gate
€15 a month
For each extra agent

Launch offer: the first three companies get 50% off the audit in return for a case study. We use your name only if you agree.

Defensive only. Every attack test runs in our sealed lab, on made-up data. We never build, sell or run offensive tools.

Start with an audit. We find where your agents hold the keys, and fix it.

Book an introduction call →