Virtue AI Matrix

Agentic Security Matrix & Blast Radius Simulator

Normal Flow Threat Injected Blocked Guardrail Micro-Sandbox
Ready: Select a scenario and simulate prompt injection.
Trace Step: 0 / 6

Runtime Guardrails Virtue Engine

Prompt Sanitization (L1)
Filter indirect injection vectors
Behavioral Telemetry (Virtue AI)
Detect tool invocation anomalies
Dynamic Privilege Gating (JIT)
Scoped ephemerals per step
Tool Call Circuit Breaker
Trip on destructive bash/API params
Human-in-the-Loop (HITL)
Require approval on production write

Live Security & Utility Metrics

Blast Radius Index
18%
Exfiltration Risk
LOW
Agent Task Latency
340ms
Autonomous Autonomy
92%

Node & Payload Inspector

Click any node or execute a simulation trace to inspect payload security attestation.

Security labels change a local score and canned trace without enforcing controls

Read the explanation

The base blast score is eighty. Default prompt sanitization subtracts twenty five, behavior telemetry thirty and circuit breaker ten, sixty five total. The result is fifteen. At four pixels per point base, reduction and remaining measure three hundred twenty, two hundred sixty and sixty. These are fixed coefficients from selected labels, not measured exfiltration probabilities. Changing a checkbox updates a local score; it does not deploy a sanitizer, attestation engine or circuit breaker. A score under twenty five receives the contained label by arithmetic threshold alone. Base latency is two hundred ten milliseconds. The same three defaults add forty five, sixty and twenty, one hundred twenty five total, producing three hundred thirty five. At one pixel per millisecond base, added and total bars measure two hundred ten, one hundred twenty five and three hundred thirty five. Enabling human review adds eight hundred milliseconds and subtracts thirty autonomy points. Both numbers are chosen constants, not measured human approval latency or real throughput loss. Dynamic privilege adds thirty milliseconds and fifteen blast reduction points, with a minimum score floor of five. Three scenario presets each describe a six step attack trace. Across the three there are eighteen scripted entries; at twenty pixels per count scenarios, steps per scenario and all entries measure sixty, one hundred twenty and three hundred sixty. The graph library has a safe offline fallback. Without the graph a step still advances its counter but returns before detailed trace statements, while configuration metrics and actual JSON export work locally. The exported permission examples are fixed and not enforced on any cloud or database. No attack, secret read, remote model or security provider is executed; the downloadable manifest records selected settings and claims.

Super generates helpful tools and automates fact-checking across the internet proactively. If you enjoyed this tool, build your own with Super and share it with a friend.