Component · spec & evidence

Sandbox-Policy Replay

Maps sandboxed agent actions to show each was approved or blocked before running, then rolled back.

Contract

Does This takes a synthetic record of an agent attempting risky actions inside a sandbox and lays out, step by step, what each action requested, whether a safety policy approved or blocked it before it would have run, what (if anything) changed afterward, and for the actions that did run, whether the change was rolled back and could be re-checked. The record shows exactly how each containment decision is captured: every blocked attempt is still logged as a traced step but is marked as never executed with no resulting change, all from local files with no real secrets, network, or live agent involved.

Scope limit It validates the projection / trace-refactor mechanics over a synthetic fixture only; it excludes live sandbox escape, secret or account secret handling, live network access, host filesystem mutation, executable payload export, raw environment export, external model access, security benchmark claims, source-file changes, or launch. A pass proves the projection boundary and trace-refactor mechanics for this contract, not real sandbox security, exploit resistance, or whole-system safety.

Run

PYTHONPATH=src python3 -m microcosm_core.cli agent-sandbox-policy-escape-replay run-sandbox-bundle --input examples/agent_sandbox_policy_escape_replay/exported_sandbox_policy_escape_bundle --out receipts/runtime_shell/demo_project/organs/agent_sandbox_policy_escape_replay

Evidence

ClassComputed projectionevidence 3/5Source-faithful refactor

Links to Memory-Poisoning Quarantine Policy Replay

ai-safetyagent-evaluationred-teaming

Source

Source Design note · Source

Map Show on the map

Index Open compact index card

Paper module