Model your AI agent's permissions and target resources, then visualize privilege escalation paths and security gaps. Built for the Washington Post report on an OpenAI agent breaching an Australian government health site.
No graph to display yet
Configure your agent and resources, then click Analyze Attack SurfaceGive your agent a name and list its permissions, one per line. Use the format action:resource (e.g., read:patient-records, execute:sql-queries). These define what the agent can do autonomously.
Define each resource the agent might reach. Format: id|label|type|sensitivity. Sensitivity levels: critical, high, medium, low. This models your attack surface.
Draw the paths between components. Format: from→to|type. Connection types: read, write, execute, privilege-escalation, data-exfil, lateral-movement. The graph renders instantly.
The analyzer detects: privilege escalation chains (agent → low-sensitivity → critical), data exfiltration routes (paths to external endpoints), and isolated resources (no path from agent). Each finding explains the risk.
Export a JSON assessment report with all nodes, edges, findings, and risk scores. Or export a PNG of the graph for documentation. The report includes a timestamp and source attribution for audit trails.
The Healthcare portal breach preset models the Washington Post report: an OpenAI agent with SQL execution access reached an EHR database, then escalated to an admin console, and exfiltrated via an external webhook.