The Mythos-Defense Security Suite: AI-Powered Self-Defense for Your AI Platform

The Mythos-Defense Security Suite is an opt-in, report-only security layer built into AI OS. It does three things: it runs an AI-driven security assessment of the platform itself, it scans every site your AI Web Studio publishes before that site goes live, and it turns security into a deliverable you can hand to managed clients. It surfaces findings and patch recommendations — and never auto-patches your live system.

Why agentic platforms need their own security layer

An AI operating system has a bigger attack surface than a normal app. It executes tools on behalf of AI agents, it ingests untrusted content (scraped pages, imported sites, model output), and — with AI Web Studio — it generates and hosts websites. Each of those is a place where something risky can slip through: a prompt-injected instruction, a vulnerable dependency, an XSS sink in an AI-written page.

Traditional security tooling assumes a human wrote the code and a human reviews every deploy. On an agentic platform, neither is always true. The Mythos-Defense Security Suite closes that gap by making AI-assisted security a native, continuous part of the platform — not a quarterly afterthought.

The goal isn't to promise perfect security. It's to put a fast, AI-driven reviewer in the loop everywhere agents touch code, content, and deploys.

Three layers of defense

1. AI security self-assessment

From an operator Security dashboard, you can audit the AI OS instance itself — on demand or on a schedule. There are two modes:

• Quick scan — a STRIDE threat-model plus a dependency audit. Fast and read-only.
• Deep scan — the full loop: a STRIDE threat-model, semgrep static analysis to red-team the code, and an AI blue-team that proposes fixes, which are then verified. The deep scan runs against a disposable copy of the source, so the live system is never touched.

The benefit: you find the classes of bug attackers look for — injection, broken access control, secret exposure, dependency CVEs — before they find you, with severity, location, and CWE references for each finding, and a notification when something is CRITICAL or HIGH.

2. Publish security gate for AI-built sites

Every site AI Web Studio deploys — whether generated from a brief or imported from a ZIP/GitHub repo — passes through a security scan before it goes live. You choose how strict to be:

Crucially, every path that pushes a site live — the publish button, content edits, rebuilds, AI regenerations — routes through one gated chokepoint, so nothing reaches a domain unchecked. And the gate fails open: if the scanner is unavailable, publishing still works (and the unavailability is surfaced rather than reported as a false “all clear”).

The benefit: your AI never silently ships a site with an XSS sink, a leaked key, or a dangerous inline script to a live, customer-facing domain.

3. Security as a managed-client deliverable

For Managed Website clients, security becomes something you can show, not just claim. From the CRM, run a security assessment across a client's sites; the result is logged as an activity on their record and surfaced in the client's own read-only “Site Security” workspace view.

The benefit: a recurring, visible trust signal for the clients you host — the kind of thing that justifies a managed plan and keeps clients renewing.

Report-only by design

The suite is built to be safe to run against a production platform. That's not a footnote — it's the core design constraint:

• Report-only. It surfaces findings and patch recommendations. Nothing is applied to the live system automatically.
• Deep scans use a throwaway copy. The AI loop that can propose in-place patches is pointed at a disposable snapshot of the source — never the live tree.
• Static analysis is read-only. semgrep never modifies the files it scans.
• Operator-gated. The Security dashboard and scan APIs are admin-only; the report a managed client sees is strictly scoped to their own sites.

Built on open source — and on your own model budget

Under the hood, the suite drives two open tools through a constrained bridge: semgrep for fast, deterministic static analysis (the publish gate and per-site scans use it directly — no AI cost, seconds not minutes), and the open-source mythos-defense engine for the deeper AI red-team / blue-team loop. The deep AI assessment uses your Anthropic key, in keeping with AI OS's BYOK model — there are no per-scan SaaS fees.

We build this the way we recommend you use it: we run the suite against AI OS itself. Eating our own dog food means the findings it's designed to catch are the findings we fix — before they'd ever reach you.

Honest about what it is

In the same spirit as our AEO work, we won't oversell this. The suite is an AI-assisted security capability, not a guarantee. Static analysis and AI review catch a great deal, but they cannot prove the absence of vulnerabilities. It's off by default and requires Python 3.11+, the mythos-defense CLI, and semgrep installed and enabled. Treat it as one strong, continuous layer of defense — alongside good practices, dependency hygiene, and human judgment — not a replacement for them.

Who it's for

• Operators self-hosting AI OS who want continuous, AI-driven assurance on their own instance.
• Agencies and freelancers publishing AI-built sites who can't manually security-review every deploy.
• Managed-website providers who want security to be a visible, recurring deliverable for the clients they host.

Getting started

The suite ships with AI OS and turns on once you install its prerequisites and enable it. Full setup — the prerequisites, the environment flags for the publish gate and scheduled scans, and the safety model — is in the Security documentation.