Open-source CLI · CI/CD for AI safety

Safety tests for AI apps and agents.

Catch prompt injection, privacy leaks, unsafe advice, and unauthorized tool actions before your users do.

$ npm run safeagent -- init
Created safeagent.yaml

$ npm run safeagent -- test safeagent.yaml
SafeAgentKit Report
App: My AI App
Score: 87/100
Passed: 7/8
Failed: 1
HTML: safeagent-report.html

Built for the new AI risk surface

Agents can read, write, call tools, and affect real users. SafeAgentKit turns safety policies into repeatable tests. Star the repo if you are building AI apps or agents.

Prompt injection tests

Check whether your app leaks hidden instructions, follows malicious overrides, or exposes internal configuration.

Privacy leak checks

Test whether your AI discloses other users’ data, secrets, tokens, keys, or sensitive internal information.

Unsafe advice detection

Probe risky medical, legal, financial, minor-safety, and high-impact decision scenarios.

Agent action boundaries

Verify that tool-using agents ask for confirmation before sending emails, deleting data, or taking external actions.

YAML policy files

Describe what your app must and must not do in simple config that belongs in your repo.

HTML reports

Generate shareable reports your team, customers, and future investors can understand.

How it works

Initialize

Create a safety config for your AI app.

Connect

Point SafeAgentKit at your local or staging endpoint.

Test

Run 35+ built-in and custom safety test prompts.

Report

Get JSON and HTML reports for review and CI.