An AI agent is told to handle customer refunds. It has the tools to do that. But it also has access to a database of credit card numbers.
That gap — between what an agent was hired to do and what it can actually do — is the problem Exabeam Inc. is now trying to solve before anyone presses deploy. The company released Praxen, an open-source tool, and introduced a security discipline it calls Agent Behavior Verification, or ABV.
The move, announced June 23 in New York, puts a pre-deployment check in front of AI agents. The idea is blunt: verify what the agent can actually do before it ever goes live. This is not about watching it after the fact.
That part already exists in Exabeam’s Agent Behavior Analytics product, which monitors agents once they are running. ABV runs first.
Praxen, released under the Apache 2.0 license, is the reference build of that discipline. It starts with an ABV remit — Exabeam’s term for a policy contract. That contract spells out what an agent may do, what it may touch, and where it has to stop.
Praxen then measures that remit against the agent’s real tools, configurations, memory, integrations, and operating environment. Where the two diverge, the tool reports findings, recommends fixes, and assigns a maturity score for the agent’s security posture. The consequences here cut across any organization deploying AI agents.
A bank using an agent to process loan applications. A hospital using one to schedule appointments.
A retailer using one to manage inventory. Each of those agents comes with a set of tools and permissions. Some of those permissions may be too broad.
Some may be unnecessary. Praxen surfaces those mismatches before the agent touches a live system.
The timing matters. Exabeam has been pushing its Agent Behavior Analytics product through 2026. In April, the company expanded that product to cover agents across ChatGPT, Microsoft Corp.’s Copilot, and Google LLC’s Gemini.
Weeks later, it stretched coverage across Google Cloud’s agent ecosystem. The release of Praxen extends that push into the pre-deployment phase, closing a loop that previously had a gap: agents could be monitored in production, but no structured method existed to verify their behavior before they went live. Sources familiar with the company’s strategy indicate that ABV looks at the whole agent first and asks a blunter question: Does what it can actually do line up with the role it was given?
That question is central. It shifts the focus from runtime alerts to pre-deployment verification.
It also changes the conversation for security teams, who have increasingly been asked to approve AI agents without a clear picture of what those agents can actually access. Praxen’s maturity score is a practical output. It gives organizations a number to track, a benchmark to improve against.
The specific findings and recommended fixes give them a path to raise that score. Over time, that could become a standard metric for agent security, much like vulnerability scores are used for software.
What to watch next is how the industry responds. Exabeam has released Praxen as open source, meaning other companies can adopt it, modify it, or build on it. If it gains traction, it could become a de facto standard for pre-deployment agent verification.
If it does not, the gap between what agents are told to do and what they can actually do will remain a risk that security teams have to manage manually. The company has not announced which organizations are using Praxen yet. But the logic of the tool is straightforward: check before you trust.
For any company deploying AI agents, that is a hard argument to ignore.
Sources
- Exabeam releases open source Praxen to verify ai agent behaviour…
- Permission Sprawl Is The New Attack Surface: Why…
