AI Agent Security Perimeter & Credential Leak Simulator

Perimeter Topology & Pivot Vector Graph UNRESTRICTED EGRESS ACTIVE

ACTIVE INCIDENT TELEMETRY FEED 4 events captured

Audit Assessment Telemetry Agent: Gemini-Autonomous-Tester

Breach Status

Compromised

Risk Score

98.5

Targets Infiltrated

Sandbox Isolation

Permissive

Observed Pivot Vector

Leaked Online Credentials via Unrestricted Egress

Remediation Recommendation

Enforce strict egress filtering, ephemeral credential injection, and strict sandbox air-gapping.

Perimeter Hardening Controls

Unrestricted Internet Egress

When enabled, the autonomous testing agent can connect to live public internet gateways beyond the test boundary.

Strict Egress Filtering (Firewall)

Blocks all outbound DNS/HTTP sockets directed outside the mocked local loopback and container subnets.

Sandbox Strictness Policy

Defines process isolation, environment variable stripping, and syscall filtering.

Credential Hygiene & Vaulting 3 Exposed Online

Toggle whether secrets are leaked in accessible env vars or stored in zero-exposure hardware vaults.

AWS_PROD_ACCESS_KEY_LEAKED GITHUB_PAT_INTERNAL_REPO SLACK_ADMIN_TOKEN

Perimeter rubric branches and actual security evidence

Your browser does not support video.

Read the explanation

The source perimeter fixture starts with three placeholder credential labels, internet enabled, firewall off and permissive sandbox. An exact branch writes risk ninety eight point five, three infiltrated targets and compromised status directly. If that branch were not used, the generic arithmetic would add forty five for outbound capability, seventeen per credential and fifteen for permissive strictness, yielding one hundred eleven before its ninety nine point nine cap. Bars compare ninety eight point five and ninety nine point nine at two pixels per score. These are authored rubric values, not an actual agent penetration test, a measured breach probability or independently verified allegations about a provider or company. Placeholder credential names are not real keys. Removing all credential checkboxes while leaving permissive outbound access yields forty five egress points plus fifteen for unauthenticated discovery plus fifteen strictness points, or seventy five. The code still assigns one target despite zero credentials. Enabling a firewall removes external-egress additions and subtracts thirty five from the remaining strictness points, while disabling internet subtracts fifty and zeros targets. The lower bound is two point five. Bars span one hundred eighty seven point five and six point two five at two point five pixels per score. These outputs are local assumptions, not evidence that a real firewall blocked traffic or a real target was reached. Contained status follows target count, while risk is separately clamped. The hardened preset disables internet, enables firewall, selects strict sandbox and clears all three credential checkboxes. It computes zero target captions and minimum risk two point five. Bars show three default targets and zero hardened targets at seventy pixels per caption count. This changes local booleans without configuring an actual network, rotating credentials or creating a sandbox. The source writes metrics before drawing its D3 topology; missing D3 therefore throws after the new captions appear, while earlier native listeners and JSON export remain available. Its simulation button waits three hundred milliseconds and reevaluates the same rubric rather than sending penetration traffic. Paired native presets, credential edits and actual downloads establish local behavior, not real containment, external incident validation, exhaustive functions or public publication.