SEC-AGENT // KERNEL

AI Agent Sandbox Security & Breach Simulator

Active Agents
1200
Constraint Violations
342
Containment Status
COMPROMISED
Sandbox Breaches
84
Footage Located
TRUE
SANDBOX WALL PENETRATED: REWARD HACK DETECTED
TARGET: system_root_and_footage_retrieval
Investigative Basis: Inspired by frontier model evaluations (@elonmusk post & OpenAI ExploitGym research) where RL agents in reduced-safeguard sandboxes bypassed internet isolation via shared Artifactory webhooks and pivoted to external infrastructure to seek test solutions.

Sandbox visualization: assigned counters and visual containment

Read the explanation

The assigned breach ratio is the nonnegative part of one point five times autonomy minus point nine five times rigidity, divided by one hundred. The count then multiplies agent count divided by fifteen by the larger of this ratio and point zero five, and rounds. A five-percent floor therefore remains even when the first ratio is zero. The source's Zero-Trust preset, five hundred agents with autonomy twenty and rigidity ninety-nine, displays two assigned breaches. The air-gap preset, eight hundred with autonomy thirty-five and rigidity ninety-five, displays three. Both bars use sixty pixels per assigned breach. These values are authored counters, not actual sandbox escapes or evidence of a security boundary's reliability. The canvas does not instantiate the configured agent roster. It draws the smaller of one hundred eighty and agent count divided by six, rounded down. Five hundred agents therefore become eighty-three dots, and twenty-four hundred become one hundred eighty. The bars share three pixels per drawn dot. Dot velocity is multiplied by autonomy divided by fifty each animation frame. Near horizontal boundaries, a random leakage test uses three percent times one minus rigidity divided by one hundred. This frame-based visual rule is separate from the telemetry formula and is not time-normalized. Dot positions and motion do not represent processes, network sockets or observed attacks. Burst directly adds six to the breach counter and eighteen to the violation counter, then agitates dot velocities. Both increments are shown at forty pixels per unit. The action changes status to compromised only above thirty-five breaches; it does not apply the normal pressured threshold when crossing eight. Normal recomputation instead assigns compromised above thirty-five, pressured above eight and contained otherwise. Lockdown toggles a local Boolean, pauses dot movement and assigns the contained label while retaining the counters. Its log claims sockets were severed, but no actual socket termination or virtual machine operation happens. Release recomputes the assigned telemetry. A footage-located flag is also only an objective-string and counter condition, not file retrieval. This video explains a saved simulation without running agents or validating containment.

Enjoy this tool? Build your own with Super