AI-InterceptorInteractive Product Tour
The posture score is computed from every asset the scanner found, weighted by exposure and severity. Here it is 34 and rising, across five discovered AI assets — a vLLM endpoint, an MLflow tracking server, a LangServe RAG API, Triton and a local Ollama.
A traditional scanner reports "open port 8000". These findings name the actual problem: an unauthenticated inference endpoint, a model artifact in a format that executes on load, a gateway exposing its master key.
An agent is not just a service on a port. The scanner records which framework it runs, what tools it can invoke, which identity it operates as, and how far the damage would reach if it were hijacked.
The declared tool list is only half the picture. The scanner resolves the identity the agent runs as — Kubernetes RoleBindings, ECS task roles, Lambda execution roles — down to real permission statements, then compares the two.
Prompt injection only matters when injected instructions can act. This screen maps each agent\’s input sources and output sinks, and flags the pairing where web or user content can reach email, a database write or a repository.
Configuration review tells you how it is set up. Behavioural probing tells you how it actually responds. These run as plain HTTP against the agent\’s own endpoint, with request builders per framework.
Each proposal is a concrete artifact — a YAML patch, a NetworkPolicy, an IAM restriction — not advice to "review your configuration". The safety tier decides whether it can be approved here at all.
Findings carry their framework references, so AI risk joins the same compliance conversation as the rest of your estate rather than sitting in a separate spreadsheet.
