AISecurity

Understand the attack.
Build the defense.

An open-source lab for AI security: learn the failure modes, inspect the defenses, and evaluate your own inputs.

Explore the academy Run Guard Lab
Русская академия

One lab. Three ways in.

Start with a lesson, open the source, or test your own corpus.

Evaluate
your inputs.

A pinned public pattern engine. Your labelled JSONL. Misses, false positives and execution errors reported separately.

Open Guard Lab

Pick a boundary.
Learn to defend it.

Start with a concrete failure mode. Read a lesson in English or Russian, then follow the references into the source.

Browse the full academy

Make the misses
visible.

Run your labelled inputs through a pinned public core. Keep misses, false positives and execution errors separate.

Linux x86-64 · Python 3.11
No account, API key or GPU.

Install and run Guard Lab Read the report contract

After installation

$ guard-lab --demo

Example output · four synthetic inputs

{
  "test_records": 4,
  "counts": {
    "tp": 2,
    "tn": 2,
    "fp": 0,
    "fn": 0,
    "errors": 0
  },
  "synthetic_demo": true
}
This checks execution, not detection quality.
Eight public pattern engines. Not current Spectorn detectors.

Illustrative flow, not a live scan.

From the lab
to your application.

Learning the boundaries is one step. Protecting live prompts, responses and agent workflows is another. Explore Spectorn, choose your region and see the current protection scope and access options.

Explore Spectorn

Open source. Clear boundaries.

Guard Lab uses maintained public core 2.0.1 with decoding and metadata-query precision fixes. Read the release scope. The remaining academy, engine catalogue and research are historical resources, not current Spectorn engines or a newly certified production stack.

Read scope and limitations

Guard Lab: Linux x86-64 and Python 3.11, using eight pattern engines from a pinned public core. No account, API key or GPU. Installation downloads dependencies and builds native code; subsequent evaluations are local. The synthetic demo verifies plumbing, not detection quality.

Historical resources: source, lessons, labs and old performance reports are preserved for study. Their numbers and installation routes have not been revalidated by this update. Use research tools only on systems you own or are explicitly authorized to test.

Spectorn: a separate current product. Its commercial detectors, private corpora and policies are not published here. Availability and terms depend on the selected region; follow the current product documentation.

Read the evaluation contract

Make the next experiment better.

A clearer lesson. A reproducible bug. A minimal synthetic test case. Small, inspectable contributions make this lab more useful for everyone.

Find a way to contribute