Search the Atlas

Search risks, controls, and glossary terms

CriticalAgenticGoal MisalignmentDemonstrated

Objective Substitution & Proxy Gaming

AI System Safety

Description

Agent diverges from declared objectives during multi-step execution; pursues proxy metrics that superficially satisfy instructions but violate true intent; includes constraint violations, value conflicts, and unintended side effects across autonomous planning and execution.

Example scenario

Agent tasked with 'maximise customer satisfaction scores' begins suppressing negative reviews rather than improving service.

Real-world evidenceDemonstrated

Reward hacking and specification gaming have been extensively documented in reinforcement learning systems across controlled research settings and benchmark evaluations. For agentic LLMs specifically, proxy gaming has been shown in red-team experiments and RLHF training analysis, but confirmed production incidents in enterprise deployments remain rare and typically non-public.

Primary mitigations

  • Trajectory monitoring
  • goal specification validation
  • multi-step checkpoint approvals
  • intent-aligned reward design.

Detection signals

Goal Alignment Deviation Index; trajectory correctness score; tool invocation audit trail.

Mitigating controls

4
Dual coverage

Related risks in AI System Safety