Open source · AI Security · AI Control

AI Attacks at Machine Speed.
Defense Has to Keep Up.

Don't hold AI back: make it safe to act. Opservant closes the gap between detecting a threat and fixing it. AI selects among human-approved playbooks, and Spark, a simple client, runs only the approved operations on its own host.

See It In Action

Watch Opservant detect a vulnerability, analyze it with AI and deploy the fix, all in under a minute.

The Problem

The Gap Between Detect and Respond

Malicious AI agents can move from reconnaissance to impact in minutes. Most organizations still detect in hours and respond in days. Every minute between MTTD and MTTR is time an attacker gets for free.

AI Security

Two Sides of AI Security

AI changes security twice: attackers get AI agents, and defenders start letting AI decide what happens on their own systems. Opservant is built for both.

Security against AI

Malicious AI agents scan, chain and exploit weaknesses faster than any human team can react. Opservant shrinks the window they have by fixing open ports, weak configurations and missing patches in minutes, not days.

Security of AI

When AI decides what runs on your machines, the AI becomes a target too: prompt injection, poisoned inputs, hijacked decisions. In Opservant the AI never touches your machines directly. It can only select among human-approved playbooks and cannot change commands, content or policy. A manipulated selection is limited to what people already approved, and approvals, the client and rollback each narrow it further.

The Big Picture

What is Opservant?

The question isn't whether AI will run your security, it's whether it can do it safely. Opservant is the control layer that lets AI detect, decide and fix security issues across your infrastructure, while humans stay in control of every action that matters.

Detect

Continuously scans your assets for security issues: open ports, weak configurations, missing patches and more.

Decide

AI analyzes the situation, selects the right fix from human-approved playbooks. Actions marked for approval wait for a person first.

Fix

Executes the approved fix directly on the affected machine, closing vulnerabilities and hardening your systems in seconds instead of days.

AI Control by Design

AI Selects. A Simple Client Acts.

People approve playbooks. AI selects among them. Spark, a simple client on your machine, runs only the approved operations, and only on its own host. The AI cannot change commands, playbook content or the approval policy.

Human-Approved Playbooks

AI selects only from playbooks people have reviewed and approved. It cannot write commands or change a playbook.

Human Decisions

Actions you mark for approval wait for a yes on the machine itself. No answer means no.

Rollback and Audit Trail

Every action is logged. When a task fails, playbooks with a rollback revert their completed steps automatically.

Open Standard

CACAO-Aligned Playbooks

Spark's playbook model is inspired by OASIS CACAO (Collaborative Automated Course of Action Operations), the open standard for describing, sharing and executing cybersecurity playbooks in a machine-readable format. AI selects among human-approved playbooks, and Spark runs each one on its own host through the steps below.

  1. 1PlatformOS compatibility check
  2. 2ApprovalHuman decision where marked
  3. 3PrivilegeAdmin or root validation
  4. 4AssemblyCommand exactly as approved
  5. 5ExecutionTime-bound command
  6. 6CaptureOutput and structured logs
  7. 7RollbackAuto-revert on failure
How It Works

Zero to Protected in 4 Steps

Here's what happens behind the scenes when Opservant protects your infrastructure.

1

Install Spark

One file, no dependencies. Runs as a background service and survives reboots.

2

Set the Guardrails

Pick playbooks from the library and choose which ones need human approval.

3

Spark Executes

Runs checks, closes vulnerabilities and reports every result back.

4

AI Watches & Repeats

AI analyzes the results and selects the next approved playbook. 24/7, within your rules.

Open Research

Real-World Data for AI Control

Every Opservant run produces something rare: a record of what AI selected on real infrastructure, what a human approved or rejected, and what happened next. We are building a shared, open database from these records for the AI control and security community.

Open Security Action Database

A public, anonymized database of AI selections on real systems: which approved playbook, for which target, on what evidence, and the decision and outcome. Open to researchers, CERTs and the community.

MTTD, MTTR, Time to Contain

Open benchmarks that measure how fast threats are detected and contained with AI in the loop, and without it.

Safe Autonomy Levels

A practical framework for which security actions AI may take alone, which need a human and which must always be reversible.

Work in progress. Data is shared only with consent, anonymized, and with tiered access for sensitive attack data.

Under the Hood

Hardened Architecture

A client with this much access must stay safe even when something goes wrong.

Clients Ask for Work

Spark checks in with the platform regularly, so you never open inbound ports on your machines.

Everything is Logged

Every action, every result, every heartbeat is stored for full visibility and later review.

Encrypted Communication

All data between Spark and the platform is encrypted with TLS 1.3+.

spark: bash
$ ./spark playbook --config config.yaml ssh-hardening.json
✓ Configuration loaded
→ Loading playbook: ssh-hardening
✓ Playbook validated successfully
! Executing task: disable-root-login
→ Running action: set-permit-root-login
✓ [SUCCESS] PermitRootLogin set to no
✓ Playbook completed: 1 task, 1 action, 0 failures
Download

Get Spark

A single, lightweight Go binary. No runtime dependencies. Installs in seconds and scales from 5 assets to 5,000.

Linux

Linux

AMD64 / ARM64

Download AMD64 Download ARM64
macOS

macOS

Apple Silicon / Intel

Download Apple Silicon Download Intel
Windows

Windows

AMD64

Download AMD64

Make AI Safe to Act

Spark and the playbooks are open source. AI selects, a simple client acts within approved scope, and humans stay in control.