A unified security event model for securing LLM, RAG, and Agent applications.

ASB Security Schema defines a canonical JSON structure for AI security events:

• 🔒 Make AI security policies easier with one standard input for OPA / Policy-as-Code
• 🧾 Standardize logs & audit trails for EU AI Act, ISO 27001, ISO 42001 and internal governance
• 🧩 Works with any LLM / RAG / Agent stack – LangChain, Dify, AutoGen, CrewAI, custom apps…

This repo is a specification repository: it contains the specification, JSON Schema, examples, and OPA policy samples.
Runtime components (such as asb-secure-gateway) use this schema as their canonical event format.

• 📄 Spec: ASB Security Event Schema v0.1
• 🧬 JSON Schema: asb-security-schema-v0.1.json
• 🧪 Examples: examples/ – LLM / RAG / Agent events
• 🧯 Policies: policies/ – OPA / Rego samples

For a Chinese overview, see README_zh.md.

Need to integrate the schema directly in your app? Use the lightweight SDKs included here.

• Python – pip install git+https://github.com/asb-security/asb-security-schema.git#subdirectory=python

  from asb_security_schema import SecurityEventBuilder, validate_event builder = SecurityEventBuilder( subject={"user": {"id": "user-123", "type": "human"}}, operation={ "category": "llm_completion", "name": "chat", "direction": "input", "model": {"name": "gpt-4o"}, }, resource={"llm": {"messages": [{"role": "user", "content": "hello"}]}}, ) event = builder.build() # validate_event(event) already runs by default

• Go – go get github.com/asb-security/asb-security-schema/go/securityschema

  import "github.com/asb-security/asb-security-schema/go/securityschema" payload := map[string]any{ "schema_version": securityschema.SchemaVersion, "event_id": "evt-123", "timestamp": "2024-01-01T00:00:00Z", "subject": map[string]any{"user": map[string]any{"id": "user-123", "type": "human"}}, "operation": map[string]any{"category": "llm_completion", "name": "chat", "direction": "input"}, "resource": map[string]any{"llm": map[string]any{"messages": []any{map[string]any{"role": "user", "content": "hello"}}}}, } if err := securityschema.Validate(payload); err != nil { panic(err) }

Maintainers: run python -m scripts.sync_schema_assets whenever the canonical schema changes to keep the SDKs in sync.

asb-security-schema is a data model for describing security-relevant actions in AI systems.

It defines:

• A common ASB Security Schema for:
  • LLM completions (chat / completion / embedding)
  • RAG (Retrieval-Augmented Generation) queries
  • Agent tool / action executions
• A set of JSON examples for typical events
• A few OPA (Open Policy Agent) policy samples that consume this schema

This repo does not implement a gateway itself.
Runtime components such as asb-secure-gateway use this schema as the canonical input for:

• Policy decisions (allow / deny / mask / escalate)
• Audit logs and forensic analysis
• Compliance and reporting (e.g., EU AI Act, internal governance)

The ASB Security Schema aims to:

1. Standardize how AI security events are represented across LLM, RAG, and Agent use cases.
2. Enable Policy-as-Code using engines like OPA, by providing a consistent input shape.
3. Make it easier to export AI security events into SIEM / observability / audit systems.
4. Support both:
   • Real-time enforcement (pre- / post-decision events)
   • Post-incident analysis (rich context for investigations).

It is a data model, not a full security product or WAF / SIEM replacement.

At the core of this schema is a single object:

SecurityEvent – a JSON document that describes one security-relevant action or decision in an AI system.

Every SecurityEvent answers the questions:

• Who did something? → subject
• What did they do? → operation
• On what resource? → resource
• In which context? → context
• With which decision and risk level? → decision (optional for pre-decision events)

All events follow this envelope:

{ "schema_version": "asb-sec-0.1", "event_id": "uuid-1234", "timestamp": "2025-01-01T12:00:00Z", "tenant_id": "tenant-a", "app_id": "kb-copilot", "env": "prod", // dev | test | prod "subject": { /* who */ }, "operation": { /* what */ }, "resource": { /* on what */ }, "context": { /* extra context */ }, "decision": { /* policy result (optional) */ } }