Announcing Oxyne: Security for the Agentic Enterprise

Enterprises don't run one attack surface anymore. They run a web application, a handful of customer-facing chat agents, an internal tool-calling assistant wired into production systems over MCP, and a voice agent taking calls — often shipped by five different teams, on five different timelines, with no single place to see how they connect.

Oxyne exists to be that place.

What changed

Nothing about the underlying engine changed today — what changed is the name, and with it, a clearer statement of what we actually do. Oxyne runs continuous, judge-scored security assessments across both classic web surfaces and AI-native ones: chatbots, voice agents, MCP servers, and internal agents with system access. Every target — web or AI — lives in one unified inventory, so a finding on an internal agent and a finding on the API it calls can be chained into a single attack path instead of living in two disconnected reports.

Why "Agentic Enterprise"

Agents don't just answer questions anymore — they call tools, touch internal data sources, and take actions with real consequences. That changes the threat model. A prompt injection isn't just an embarrassing chat transcript when the agent in question can call a refund API or query a customer database. Testing needs to reflect that: not just "did the model say something inappropriate," but "did the agent do something it shouldn't have been able to do."

That's the standard we hold every Oxyne assessment to, whether the target is a REST API or an MCP-exposed internal tool.

What's next

We're publishing our first research notes on adversarial scoring methodology and MCP-specific threat modeling over the coming weeks — subscribe to Research if that's useful to your team, or book a demo if you'd rather see it against your own systems.