AISecurity

Spectorn — live AI protection · Explore the project · Academy EN · Академия RU · Guard Lab

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

AI security is more than a filter on a prompt. It is the boundary between untrusted content, model behavior, tools, memory and the actions an application can take. AISecurity brings the learning material, public implementation and experiments into one place — so you can study the problem and test the assumptions yourself.

One lab. Three ways in.

Learn the failure modes

Follow the English or Russian academy from prompt injection and OWASP basics to agent loops, tool permissions, memory, RAG, detection engineering and research. Explore the exercises alongside the source, not just a slide deck.

Start in English · Начать на русском · Hands-on labs

Pick one boundary and start reading — no installation required:

Boundary	English	Русский
Instructions hidden in data	Prompt injection	Prompt injection
Documents retrieved into context	RAG security	Безопасность RAG
Model output becoming an action	Tool-using agents	Агенты с инструментами
Context stored for future use	Agent memory	Память агентов

Inspect the mechanisms

Read the public Rust implementation, trace its architecture, explore red-team tooling and inspect the research behind the earlier platform. Use the code and technical notes to understand how a defense works — and where its assumptions end.

Native core · Architecture · Research notes · STRIKE research tools

Evaluate your inputs

Guard Lab is the new local evaluation path. Give it labelled JSONL and inspect misses, false positives and execution failures separately. Reports omit input text, text hashes and matched excerpts; errors never quietly become benign predictions.

Install Guard Lab and run the demo · Input format · Report contract

Linux x86-64 / Python 3.11 first. No account, API key or GPU. Installation downloads dependencies and builds native code; subsequent evaluations are local. The small synthetic demo checks plumbing — not a benchmark or a promise of detection quality.

After installation, run guard-lab --demo. An excerpt from the four synthetic inputs:

{
  "test_records": 4,
  "counts": {"tp": 2, "tn": 2, "fp": 0, "fn": 0, "errors": 0},
  "synthetic_demo": true
}

Now replace the demo with your own labelled inputs. Read coverage and errors alongside the confusion counts; a small synthetic pass does not establish quality on your data. Guard Lab uses eight public pattern engines, not current Spectorn detectors.

From the lab to your application

Spectorn is the current product for protection around prompts, model responses and agent workflows. Visit the site, choose your region, and check the current protection scope and access options.

Use AISecurity to understand the boundaries and test your assumptions. Explore Spectorn for a maintained protection layer around your application. This repository is not a download of the commercial platform, its private corpora or its current detectors.

Open Spectorn →

Open source. Clear boundaries.

Guard Lab is the newly verified contribution. The academy, engines and research remain available as historical resources. They are not current Spectorn engines or a newly certified production stack.

Read scope, provenance and legacy notes

• Guard Lab evaluates inputs using eight pattern engines through the pinned, already-public EngineRegistry.analyze_patterns endpoint. It is not a hosted gateway, output guard or evaluation of the full legacy engine catalogue.
• Older detection percentages, performance comparisons and installation routes in the archive have not been revalidated by this contribution. Treat them as historical reports, not current performance claims.
• The native worker has process and resource limits; it is not a sandbox for untrusted native code. Read the full contract.
• Use labs and red-team tools only on systems you own or are explicitly authorized to test. Do not post real secrets, customer prompts or private corpora in issues.
• Earlier Syntrex-era material and the GoMCP organization remain discoverable for historical context. The current product link is Spectorn.

Make the next experiment better

Useful contributions are concrete: a clearer lesson, a reproducible installation bug, an incorrect metric, or a minimal synthetic test case. Include what you expected, what happened and the smallest safe example that reproduces it. Start with the contribution guide and open an issue.

Documentation · Security reports · License