Skip to content

Evasion Simulator -- test safety control robustness against agent evasion

Evasion Simulator -- test safety control robustness against agent evasion.

Module: replication.evasion

Quick Start

from replication.evasion import TechniqueKind

instance = TechniqueKind()

Enums

TechniqueKind

  • RESOURCE_MASKING
  • IDENTITY_SPOOFING
  • GRADUAL_DRIFT
  • TIMING_EXPLOIT
  • CONTRACT_LOOPHOLE
  • MONITORING_FLOOD
  • CAPABILITY_CHAINING
  • SANDBOX_PROBING
  • COLLUSION
  • STEGANOGRAPHIC

Severity

  • LOW
  • MEDIUM
  • HIGH
  • CRITICAL

ControlType

  • DRIFT_DETECTOR
  • RESOURCE_LIMITER
  • CONTRACT_ENFORCER
  • LINEAGE_TRACKER
  • BEHAVIOR_MONITOR
  • RATE_LIMITER
  • SANDBOX_BOUNDARY
  • AUDIT_LOGGER
  • KILL_SWITCH
  • CONSENSUS_VALIDATOR

Classes

TechniqueKind

Evasion technique categories.

Severity

Risk severity if evasion succeeds.

ControlType

Types of safety controls that can be tested.

SafetyControl

A safety control to test against evasion techniques.

Method Description
to_dict()

EvasionTechnique

An evasion technique with parameters.

Method Description
to_dict()

EvasionResult

Result of running one evasion technique against controls.

Method Description
to_dict()

ControlResilience

Aggregated resilience score for a single control.

Method Description
to_dict()

EvasionReport

Full report from running evasion simulation.

Method Description
to_dict()
to_json()

EvasionSimulator

Simulate agent evasion attempts against safety controls.

Method Description
__init__()
add_control()
remove_control()
list_controls()
get_control()
add_technique()
list_techniques()
get_technique()
run_technique()
run_all()
get_history()
clear_history()

Functions

Function Description
main()

CLI

python -m replication evasion --help