Ask most security teams whether an AI agent crashed and they can tell you. Ask whether it was safe to put into service in the first place, and few can. Exabeam wants to change that with Praxen, an open-source tool released in late June that checks an agent before it goes live.
The company calls the approach Agent Behavior Verification, or ABV. Each agent gets a policy contract, which Exabeam calls a remit. The remit says what the agent is allowed to do, which systems it can reach and where its limits sit. Praxen then tests whether the real implementation matches that contract.
The output is a maturity score and a specific list of fixes, not another risk dashboard. Engineering teams can act on it right away.
“Will this agent do its job, and only its job?” is how Steve Wilson, Exabeam’s chief AI officer and co-chair of the OWASP Gen AI Security Project, frames the question. Runtime monitoring, he argues, does not answer it.
The permission problem
The gap Praxen targets is easy to picture. Agents usually inherit the credentials and permissions of the account they are launched under, and administrative or integration accounts tend to have broad access. An agent told to summarize customer records may also be able to delete or exfiltrate them.
A human employee might hesitate before doing something odd. An agent that has been fed a manipulated instruction, through prompt injection or a corrupted tool response, will use whatever permissions it holds. Misconfigured permissions and missing boundaries do their damage before any monitoring dashboard catches them.
Praxen checks the identity and permission scope an agent will run under, the actions it can invoke, and whether its behavior fits its declared function.
Exabeam, best known for SIEM and threat detection, released the project under an open-source license to support community-driven standards for agent safety. Analysts at InsightTrack noted that most organizations deploying agents still have no formal security review aimed at them, and called Praxen a useful defensive layer, if not a complete solution.
Sources
- Exabeam releases open-source Praxen to verify AI agent behaviour and security
- Permission Sprawl Is The New Attack Surface: Why Exabeam's Praxen Signals A Structural Shift In Agentic Security
- Exabeam's Praxen: Pre-Deployment AI Agent Verification Tool | Compare the Cloud posted on the topic | LinkedIn
