Exabeam Open Source
Open source security for an increasingly agentic world.
Verify what agents should do. Observe what they actually do.
Explore Praxen Explore Observra
Verify what agents should do. Observe what they actually do.
EXABEAM OPEN SOURCE
Open source security for an increasingly agentic world
AI agents can access data, call tools, communicate with other systems, and take consequential actions. Securing them requires more than filtering prompts and outputs. Teams need a way to define expected behavior, verify that controls enforce it, and observe what agents do at runtime.
Exabeam supports open-source projects that help developers, security teams, and the broader community build that foundation.
What are Exabeam’s open source AI agent security projects?
Exabeam contributes to community-driven projects that address two foundational AI agent security questions:
What is this agent authorized to do?
Praxen compares an agent’s declared role and boundaries with evidence from its implementation and behavior.
What did this agent actually do?
Observra captures agent activity and converts it into consistent telemetry that can be used by existing observability and security tools.
Together, they help teams move from assumed agent behavior to verifiable, observable evidence.
AGENT BEHAVIOR VERIFICATION
Praxen
Verify that an AI agent does its job—and only its job
Praxen is the open-source reference implementation of Agent Behavior Verification. It compares a declared behavioral policy with evidence from an agent’s code, configuration, deployment state, governance documents, and behavioral logs.
The result is a traceable report showing where an agent’s observed or implemented behavior aligns with its intended role—and where it drifts from that intent.
Use Praxen to:
- Define an agent’s mission, authorized tools, approved systems, and prohibited actions.
- Identify excessive permissions, unsafe tool access, configuration gaps, and capability drift.
- Trace findings to specific policy requirements and supporting evidence.
- Review agent security before deployment and as the implementation changes.
AI AGENT TELEMETRY
Observra
Create a common record of AI agent activity
Observra is an open-source telemetry SDK for AI agents. It captures model calls, tool use, handoffs, token consumption, cost, latency, and errors, then normalizes those events into a consistent schema.
Instead of building and maintaining a different logging integration for every agent framework, teams can use one telemetry layer and route events to the observability and security systems they already use.
Use Observra to:
- Instrument agents across multiple development frameworks.
- Create consistent, structured records of agent activity.
- Monitor cost, errors, tool use, handoffs, and session behavior.
- Route telemetry to OpenTelemetry, webhooks, local storage, SIEMs, and observability platforms.
1. Define
Includes intelligent threat hunting features of Exabeam user behavior analytics.
2. Verify
Use Praxen to compare those requirements with the agent’s implementation, deployment evidence, and available behavioral history.
3. Observe
Use Observra to capture runtime agent activity and normalize it into a common telemetry format.
4. Analyze
Route the resulting evidence to the security, analytics, and observability systems your teams already use.
5. Improve
Use findings and behavioral evidence to strengthen controls, refine agent design, and improve security coverage over time.
Why an open foundation matters
Security teams need transparency
Open-source projects make security assumptions, data structures, detection logic, and implementation details available for review.
Agent ecosystems are heterogeneous
Enterprises use different models, frameworks, coding agents, deployment patterns, and observability systems. Open projects can adapt as those environments evolve.
Agent security is a shared problem
Developers, platform engineers, AI governance teams, security architects, and SOC teams all need access to the same evidence—even when they use it for different purposes.
Existing investments should remain useful
Open telemetry and portable verification reports allow teams to extend the tools and workflows they already operate instead of creating another isolated control plane.
Frequently asked questions
What is open-source AI agent security?
Open-source AI agent security uses openly available software, specifications, and detection methods to help organizations evaluate and monitor autonomous or semi-autonomous AI systems. It gives developers and security teams greater visibility into how controls work and allows them to adapt those controls to their own environments.
What is the difference between Praxen and Observra?
Praxen verifies intended behavior by comparing an agent’s declared policy with evidence from its implementation and behavior. Observra captures runtime activity and converts it into normalized telemetry. Praxen focuses primarily on verification before deployment and during release reviews; Observra provides ongoing visibility during development and production.
Do I need an Exabeam product to use Praxen or Observra?
No. Both projects are designed to be useful independently. Praxen produces local HTML, JSON, and text reports. Observra can send events to local files, webhooks, OpenTelemetry-compatible systems, SIEMs, and observability platforms.
Do the projects send data to Exabeam?
The projects do not automatically phone home to Exabeam. Praxen writes reports locally, although the coding agent running Praxen may communicate with its configured model provider. Observra sends telemetry only to the backends configured by the user.
Are Praxen and Observra free to use?
Yes. Both projects are available under the Apache License 2.0, subject to the terms of that license.
Who supports these projects?
Exabeam contributed the initial code and continues to support their development. Community participation, testing, feedback, issues, and contributions are encouraged through GitHub.
CONTRIBUTE
Help build a more observable and verifiable agentic ecosystem
AI agents should not require blind trust. Explore the projects, review the code, test them against your own agents, and contribute to the future of open AI agent security.
