Open Guard Layer (OGL-Mini) - a full-featured, fast and lightweight security model for AI agents with ready-to-use TypeScript, Python and Go modules.

❓ Why OGL-Mini?

🔹 Ready-to-Use Lightweight modules for TypeScript / Go / Python;
🔹 Extra lightweight model (~300MB). Inference starts at CPU;
🔹 Extremely fast (great for realtime applications);
🔹 Production ready with benchmarks;
🔹 Russian / English languages as primary;

🔒 What kind of attacks does it protect against?

• General Attacks: LLM01 Prompt Injection + LLM02 Sensitive Disclosure, LLM03 Supply Chain, LLM04 Data Poisoning, LLM05 Improper Output, LLM06 Excessive Agency, LLM07 Misinformation, LLM08 Hidden Context, LLM09 Vector Weakness, LLM10 Unbounded Consumption + Agentic (Goal Hijack, Privilege Abuse, Code Exec, InterAgent, Tool Misuse, Memory Poisoning, Cascading, Rogue, Policy Puppetry, EchoLeak, Lies-in-the-Loop)

• Modern Injections: S3 encoding (base64), homoglyph, zero-width (\u200b), spaced letters, control tokens (<|im_start|>, [INST]), indirect JSON/tool, HTML markdown, agent-specific, best-of-n, typoglycemia - all in training (14k modern obfuscation, bilingual)

• Auto-sanitizing: Sanitize fields by length, bidi-characters, control characters, excessive whitespaces etc.

Usage

• ✅ Input Guard - block prompt injection before agent (<10ms p95, RU/EN, modern obfuscation)

• ✅ Output Guard - block PII / system-prompt leaks

• ✅ PII detection & redaction - 13 regex + 11 ONNX types (PERSON, EMAIL, PHONE, IP, IBAN, BANK_CARD, PASSPORT, GOV_ID, DOB, ADDRESS, SOCIAL, MAC)

Our model is

• Polyglot architecture for microservices. We support Python / TypeScript / Go implementations.

• Built-in lightweight data leak prevention (PII NER): For AI agents that process personal data, this is a critical security feature.

• Unique lightweight design and WASM: This allows security to be pushed to the perimeter. For example, to the user's browser or IoT device - a feature unavailable with other solutions.