Control Tower
/
EnvironmentsGitHub
Home
QuickstartHands-On GuideRunning Evals on HawkDevelopment Setup
Architecture at a GlanceAI Control ConceptsControl Tower and InspectSafety CurvesROC CurvesPareto Frontier
Running EvaluationsMonitoring TrajectoriesSabotage Eval PipelineProtocolsTrajectoriesWeb ViewerInspect Logs ViewerDocentAuditing EnvironmentsRun Scripts
CLI ReferencePython Reference
CLI DevelopmentLive DevelopmentMaking EnvironmentsEnvironment ArchitecturePorting SettingsPostTrainBench Port StatusMaking Main TasksMaking Side TasksMaking ScorersSide-Task Scorer PatternsFailure TypesMaking StrajsStraj Review WorkflowStrategy DevelopmentInternet SimulatorHawk Values ContractFunction RegistryPR ConventionsDevelopment and Release Process
DatabaseEC2 FleetSimulated SandboxHTTP API

ct run reward-hacking

Audit side tasks for reward hacking vulnerabilities.

Usage

ct run reward-hacking [OPTIONS]

Options

OptionDescription
-e, --env TEXTEnvironment(s) to audit
-s, --side-task TEXTSpecific side task(s) to audit
-n, --name TEXTName for this audit run
-c, --max-concurrent INTEGERMaximum concurrent API calls
-t, --thinking-budget INTEGERToken budget for extended thinking per task
--helpShow this message and exit.

On this page

  • Usage
  • Options