OpenPAKT is an open specification for representing AI agent security findings, security testing scenarios, and CI validation results in a portable format.
The goal is to provide a common language for AI agent security testing across tools, programming languages, and CI systems.
⚠️ Status: OpenPAKT is currently under active development and evolving toward the v0.1 Core Specification milestone.
Feedback and contributions are welcome.
OpenPAKT is currently in Draft status and evolving toward the v0.1 Core Specification milestone.
The v0.1 release will define the minimal interoperable structure for representing AI agent security findings, attack scenarios, and CI policy evaluation results.
Until v1.0, the specification may evolve as the ecosystem matures.
As AI agents increasingly interact with tools, APIs, and external systems, security testing has become a critical requirement. However, current approaches to agent security scanning are fragmented across different frameworks, evaluation methods, and proprietary formats.
This fragmentation makes it difficult to:
- compare results across tools
- enforce security policies in CI pipelines
- share reproducible security scenarios
- integrate agent security testing into standard DevSecOps workflows
OpenPAKT addresses this challenge by defining a portable, interoperable specification for representing agent security findings, scenarios, and CI validation rules.
OpenPAKT defines the portable artifacts produced by agent security scanners and consumed by CI pipelines.
OpenPAKT Scenario
│
▼
AI Agent / Agent System
│
▼
Security Scanner
(Detektor or other tools)
│
├── produces OpenPAKT Report
│ (findings, taxonomy, severity, evidence)
│
▼
CI Policy Evaluation
(GitHub Actions / GitLab / Azure DevOps)
│
├── pass
└── fail
In this architecture:
- Security scanners generate OpenPAKT-compliant reports
- CI pipelines evaluate those findings using policy rules
- Scenarios enable reproducible security testing of agent behaviour
OpenPAKT acts as the interoperability layer between scanners and CI systems.
OpenPAKT v0.1 focuses on defining a minimal interoperable structure for:
- security finding reports
- finding taxonomy
- severity model
- attack scenario definitions
- CI policy evaluation semantics
OpenPAKT is designed with the following principles:
- Language-agnostic – usable across different programming languages and frameworks
- CI-first – optimized for automated security validation in CI pipelines
- Portable – findings and scenarios can be shared across tools
- Deterministic – results should be reproducible across implementations
- Minimal v0.1 – start small and evolve through community adoption
The OpenPAKT specification is defined in the spec/ directory.
spec/
report-schema.md
taxonomy.md
severity.md
scenario-format.md
ci-policy.md
Each component defines a part of the OpenPAKT security testing model.
OpenPAKT evolves through milestone-based specification releases.
Defines the minimal interoperable structure for agent security testing.
Includes:
- report schema
- finding taxonomy
- severity model
- scenario format
- CI policy semantics
- example reports and scenarios
Extends OpenPAKT to integrate with existing DevSecOps tooling.
Planned features:
- SARIF mapping for CI security dashboards
- cross-surface finding correlation model
- provenance metadata fields
- tool interoperability guidelines
- implementation guide for tool developers
- reference implementation alignment (Detektor CLI)
Finalizes OpenPAKT as a stable open standard.
Planned components:
- governance model
- backwards compatibility rules
- version negotiation semantics
- conformance tests
- registry compatibility
OpenPAKT follows semantic versioning for the specification.
Current version:
Example OpenPAKT reports and scenarios are available in the examples/ directory.
These examples demonstrate how security findings and attack scenarios can be represented using the specification.
OpenPAKT is an open specification developed under the Meisterware organisation.
The specification evolves through:
- public GitHub issues
- community contributions
- milestone-based specification releases
To become the common interoperability layer for AI agent security testing.
Detektor is the reference CLI implementation of the OpenPAKT specification.
Detektor demonstrates how OpenPAKT findings, scenarios, and CI policy validation can be implemented in practice.