AI & Model-Level Risks

Model Manipulation / Prompt Injection

Attack: Submitting signals crafted to exploit AI behavior or bias models.

Mitigations:

• Signals are schema-validated

• No free-form text inputs

• Deterministic encoding of human inputs

• Multiple independent agent evaluations

Residual Risk: Low—attack surface intentionally constrained.

Model Drift & Overfitting

Risk (Non-Adversarial): AI agents overfit to recent regimes or degrade in performance.

Mitigations:

• Continuous regime detection

• Ensemble agent disagreement thresholds

• Auto-deployment of fallback strategies

• Performance-based agent throttling

Residual Risk: Moderate, but actively managed.

AI-Agent Collusion

Attack (Theoretical): Multiple agents converge on a flawed strategy due to shared biases.

Mitigations:

• Heterogeneous agent architectures

• Independent training pipelines

• Forced diversity constraints

• Capital allocation caps per strategy