OSSASAI - Open Security Standard for Agentic Systems

A vendor-neutral, community-driven security framework designed specifically for AI agent systems that interact with external tools, filesystems, networks, and users.


Overview

OSSASAI provides a structured approach to securing AI agent systems against unique threats including prompt injection attacks, tool abuse, context poisoning, identity confusion, and capability escalation.

Design Philosophy: “Access control before intelligence.” Most AI agent security failures are not sophisticated exploits—they’re cases where “someone messaged the bot and the bot did what they asked.” OSSASAI’s stance: Identity first, scope next, model last.


Framework Structure

Section Description
Specification Core requirements using RFC 2119 normative language (MUST/SHOULD/MAY)
Threat Model Adversary taxonomy, attack vectors, and AI-specific threats
Controls 30 security controls across 8 domains
Implementation Deployment guides for each assurance level
Testing Verification procedures and automated auditing
Compliance Evidence collection and reporting
Incident Response Playbooks and recovery procedures
Profiles Platform-specific implementation mappings

Assurance Levels

OSSASAI defines three assurance levels based on deployment context and risk tolerance:

Level Name Description
L1 Local-First Single-user, loopback-only deployments with minimal attack surface
L2 Network-Aware Multi-user deployments with LAN/VPN exposure
L3 High-Assurance Production deployments with public exposure or regulatory requirements

Trust Boundaries

The framework defines four trust boundaries that must be secured:

Boundary Name Description
B1 Inbound Identity Message sources and sender verification
B2 Control Plane Administrative interfaces and configuration
B3 Tool Governance Capability restrictions and sandboxing
B4 Local State Secrets, logs, and persistent data

Control Domains

Domain ID Controls Description
General GEN 5 Security by default, fail secure, least privilege, defense in depth, audit logging
Control Plane CP 4 Gateway exposure, authentication, proxy trust
Identity & Session ID 3 Peer verification, session isolation, group policies
Tool Blast Radius TB 4 Least privilege, approval gates, sandboxing
Local State LS 4 Secrets protection, log redaction, retention
Supply Chain SC 3 Plugin trust, reproducible builds, artifact signing
Formal Verification FV 3 Security invariants, testing, CI integration
Network Security NS 4 TLS, certificates, API security

Getting Started

  1. Understand the threat model - Review adversary classes and AI agent threats
  2. Choose your assurance level - Based on deployment context, select L1, L2, or L3
  3. Review applicable controls - Identify controls required for your assurance level
  4. Implement controls - Follow implementation guides or use a platform-specific profile
  5. Verify compliance - Use automated auditing and evidence collection

Implementation Profiles

OSSASAI is designed to be implemented by specific AI agent platforms through implementation profiles. Each profile maps OSSASAI controls to platform-specific features, configurations, and tooling.


Standards Alignment

OSSASAI maps to established security frameworks:

Standard Alignment
OWASP ASVS v4.0 Authentication, session management, access control
NIST SP 800-53 AC, AU, CM, IA, SC control families
NIST AI RMF AI-specific risk management
CIS Controls v8 Controls 3, 4, 5, 6, 12, 16
MITRE ATT&CK Tactic and technique mapping

Contributing

OSSASAI is an open standard. Contributions are welcome:

  • Framework improvements - Submit issues and PRs to the OSSASAI repository
  • Implementation profiles - Create profiles for additional AI agent platforms
  • Security research - Report vulnerabilities and contribute threat intelligence

Version

OSSASAI v0.2.0 - January 2026

This release includes 30 security controls, a complete profile registry system, and production-ready tooling.


Back to top

OSSASAI v0.2.0 - Open Security Standard for Agentic Systems. Apache 2.0 License.

This site uses Just the Docs, a documentation theme for Jekyll.