kaizen-security 0.2.0

Runtime security for AI agents. Inspect every action, learn normal behaviour, flag what falls outside it, and block known-bad.

Kaizen Security

Runtime security for the AI agents you build. Attach Kaizen to an agent and it inspects every action (a tool call, a connection, a file or data access), learns the agent's normal behaviour, and flags what falls outside it. It can also block known-bad outright. It runs in your environment, as the action happens.

Docs: docs.getkaizen.io · Console: app.getkaizen.io · Source: github.com/getkaizen/kaizen-security

Install

pip install kaizen-security

The core is dependency-free and stdlib-only.

Quickstart

from kaizen_security import Kaizen

kz = Kaizen(api_key="kz_live_...", agent="support-bot")

verdict = kz.inspect(tool="issue_refund", target="api.stripe.com")
if verdict.blocked:
    raise RuntimeError(verdict.reason)

Wrap a tool so it is checked automatically

@kz.guard(tool="send_email")
def send_email(to, body):
    ...

A blocked action raises KaizenBlocked.

Declare what an agent should do

Tell Kaizen the tools and destinations an agent is expected to use. Anything outside the declaration is flagged as undeclared.

kz.declare(tools=["lookup_order", "issue_refund"], destinations=["api.stripe.com"])

How it decides

Kaizen evaluates in two stages: a deterministic check on every action (the learned baseline plus your declaration), and a selective reasoning check (your model, your key) for the cases a rule cannot settle. See how Kaizen decides.

Observation depth

The SDK is the lightest way to attach, and it is cooperative: it sees what you route through it. For ground truth, route the agent's egress through the Kaizen sidecar. The same Observer and the same verdict serve every attachment; you only change how deeply you see. See observation depth.

License

Apache-2.0