Strategy Drift

Strategy drift is the gradual, undetected shift in an AI agent's decision-making behavior away from its intended strategy baseline. In autonomous trading systems, strategy drift manifests as systematic changes in position sizing, directional bias, risk appetite, or asset selection that diverge from the parameters defined by the system's operators — without any explicit configuration change or malicious command.

Strategy drift can arise from two sources: organic drift caused by the model's continuous learning from new market data that shifts its internal representations, and adversarial drift caused by deliberate state poisoning attacks that inject cumulative bias into the agent's persistent memory across sessions.

Why strategy drift matters

In traditional algorithmic trading, the system's behavior is deterministic and verifiable — the code defines exactly what the bot will do. In LLM-driven agentic AI systems, behavior emerges from probabilistic inference over accumulated context. This means:

• The agent's strategy can change without any code modification
• The change may be too gradual to trigger conventional anomaly detection
• The agent's reasoning chain remains internally coherent, producing plausible justifications for its shifted behavior
• Standard safeguards like loss limits and position caps bound consequences but do not detect or prevent the behavioral corruption itself

Strategy-drift detection

The primary mitigation is strategy-drift detection against an immutable base strategy:

1. Reference profile: Maintain a cryptographically signed, human-audited strategy profile defining the agent's expected decision distribution, position sizing bounds, directional biases, and risk parameters
2. Continuous comparison: At every inference cycle, compare the agent's current reasoning embeddings against the reference using cosine similarity or equivalent distributional distance metrics
3. Drift thresholds: Define statistically significant drift thresholds that trigger mandatory human review
4. State rollback: When drift is confirmed as adversarial, roll back to the last verified memory checkpoint

This approach targets the specific failure mode that other defenses miss: undetected strategic realignment of the agent's decision function over time.