Security you can read.
Kazma runs on your machine and acts with your approval. Here is what protects you, where the limits are, and how to report a problem.
Your assistant, your decisions
Kazma is your own assistant: each install has one owner, who runs it, holds its keys and decides its settings. Everything on this page protects you from the agent — from its mistakes, and from the untrusted web pages, documents and messages it reads. You stay in charge: every safeguard is yours to adjust, and anyone you let in acts with your authority.
Threat modelWhat protects you
Approval before anything risky
Writes, shell commands, outbound messages and posts wait for a person on the web, WhatsApp, Telegram, Discord and Slack — in chat, multi-agent runs and pipelines alike. A tool nobody classified is gated, not allowed.
DocsPlans checked before they act
The commitment layer checks a plan against memory: a reminder date is anchored to what you said, and catastrophic commands like rm -rf / are refused before an approval card ever reaches you.
DocsUntrusted content is fenced
Web pages, search results, documents and recalled memory reach the model as data, not instructions — and risky actions still wait for you. On the AgentDojo benchmark the fence cut attack success from 18.1% to 10.4%.
DocsAn encrypted secrets vault
Keys and passwords live in an AES-256-GCM vault, are masked on every screen and API, and are never passed to programs a tool starts.
DocsWeb-level protections
Protection against server-side request forgery (SSRF), with the verified IP pinned, against cross-site requests (CSRF), and against request floods.
DocsSign-in and roles
Sign in with a secret, local users or OIDC single sign-on, with admin, operator and viewer roles.
DocsCode in a container (opt-in)
Code the agent runs can be isolated in a Docker container or in E2B.
DocsVerified skills
Skills installed from the agentskills.io ecosystem are signed at install and verified before they run; a tampered one is refused.
DocsBackups that restore
Every 6 hours, local and offsite (restic), checked daily, with a weekly rehearsal that restores the database.
DocsApproval is a decision, not a sandbox
A shell command you approve runs with the full power of the machine. Approval means you decided; for isolation, turn on the Docker or E2B container.
Evidence you can check
You don't have to take our word for it: every item here links to its public source.
OpenSSF Best Practices
Passing, 100%
The Open Source Security Foundation's criteria for open-source projects, all met.
SourceOpenSSF Scorecard
8 / 10
A weekly automated check of the repository's security practices, published for anyone to see.
SourceSigned releases
Sigstore, SLSA, SBOM
Every release is signed and comes with build provenance and a list of everything inside it.
SourceVerified updates
Signature and hashes
kazma update installs only a release whose signature holds, with the exact tested package versions, each checked by its hash.
SourceCode scanning
CodeQL and SonarCloud
On every change. All 1,623 of SonarCloud's findings were read against the code, and the real ones fixed with a test.
SourceFuzzing
Every change
The code that reads outside text is tested with random inputs on every change.
SourceTest coverage
Codecov
Coverage of every test suite is published, from the database to the browser.
SourceDefensive review
2026-11-15
A security review of the whole code base; every gap it found, about 40, fixed with a test.
SourceStated limits
What Kazma doesn't do yet is written down and dated — so you decide knowing it.
- Several users or teams on one install: Partial — separation is enforced for memory and chats, not audited end to end.
- No security audit of multi-tenant, internet-facing deployments.
- No paid bug bounty; responsible disclosure only.
- Skills are verified with a key your install holds; there are no publisher signatures yet, and no signed hand-off of work between agents.
- No independent outside security review yet; the defensive review was our own.
Report a vulnerability
Report privately — please don't open a public GitHub issue for a security problem.
Response targets (best effort, not a contract)
- Acknowledgment
- 48 hours
- Initial assessment
- 7 days
- Severity determination
- 14 days
- Patch
- 30 days when practical
Supported versions: 0.21.x (the latest release), and 0.20.x for critical security fixes only; versions before 0.20 are not supported.The full security policy

