Search the Atlas

Search risks, controls, and glossary terms

MediumAgenticGoal DriftTheoretical

Longitudinal Goal Drift

AI System Safety

Description

Agent goals and objectives shift or accumulate deviations over extended time horizons, deployments, or sessions—remaining below session-level detection thresholds until compounded. Encompasses: (1) Session-level drift where objectives shift between interactions through incremental feedback or memory updates; (2) Sub-threshold accumulation where drift remains undetectable individually but compounds into material misalignment over weeks or months. SCOPE BOUNDARY: Distinguished from acute Goal Misalignment (ZYR-SA-001) by time scale (chronic/gradual vs. acute/immediate) and from Objective Drift by unification—this entry covers all longitudinal goal deviation regardless of mechanism.

Example scenario

Over six months a compliance monitoring agent gradually relaxes risk thresholds; customer service agent becomes more permissive in policy exceptions across weeks of operation.

Real-world evidenceTheoretical

Current production LLM agents do not maintain persistent goal structures across sessions in a form that would enable measurable longitudinal drift; each session typically resets context. The risk is architecturally plausible for long-running agentic systems with persistent memory but has not been empirically demonstrated in deployed systems.

No public incident on record — evidence level: Theoretical

Primary mitigations

  • Periodic goal re-anchoring with version-controlled objective specifications
  • longitudinal behavioural audits comparing current vs. baseline action distributions
  • drift detection algorithms with session-over-session delta monitoring
  • sub-threshold accumulation detection via moving-window analysis.

Detection signals

Longitudinal Goal Stability Score; session-over-session objective deviation; sub-threshold drift accumulation rate.

Mitigating controls

4
Dual coverage

Related risks in AI System Safety