Guardrails

[All Actors]
1. Identify critical guardrail requirements, by reviewing applicable regulatory standards (e.g., GDPR) and internal policies and by assessing potential impact on safety, privacy, fairness and transparency. 

2. Implement Technical Safeguards:
a. Develop Input and Output Validation Mechanisms:
- Ensure the AI model processes only authorized inputs and generates outputs within pre-defined acceptable parameters.
- Apply filters to flag or reject anomalous inputs or outputs, particularly in cases where outputs could be harmful.
- Leverage canary tokens to detect potential prompt leakage attacks and block responses containing unauthorized system instructions.
b. Set Thresholds for Model Confidence:
- Establish a threshold confidence level below which the AI system must defer to human intervention or request additional validation.
- Log instances when outputs fall below the threshold for future analysis and system improvement.
c. Integrate Risk-Based Guardrails:
- Apply risk-scoring methods to outputs, adjusting system responses accordingly based on risk level (e.g., higher scrutiny for high-impact decisions).

3. Develop Fail-Safe Mechanisms:
a. Implement Emergency Stop Mechanisms:
- Include a shutdown feature that can be triggered automatically in response to detected anomalies, or manually by a human supervisor.
- Ensure that shutdowns do not lead to data loss and create recovery checkpoints where possible.
b. Set Up Recovery Protocols:
- Establish recovery protocols that enable safe system restart following a shutdown, ensuring that unresolved issues are addressed before the system is reactivated.

4. Establish Governance and Monitoring Procedures:
a. Assign a Guardrail Review Team:
- Designate a team responsible for reviewing guardrail effectiveness, responding to flagged issues, and recommending guardrail adjustments based on system performance and risk assessment.
- Ensure the team includes personnel capable of intervening and making decisions in case of system failures or unexpected outputs.
b. Continuously Monitor System Performance and Risk Indicators:
- Implement continuous monitoring tools to track system performance, accuracy, and risk indicators in real time.
- Set up automated alerts to notify the guardrail review team of anomalies or potential risks.
- Include human-in-the-loop (HITL) mechanisms for reviewing flagged anomalies or when the system reaches critical decision thresholds.

5. Define Operational Protocols for Guardrail Adjustments:
a. Regularly Review and Update Guardrails:
- Conduct regular reviews of guardrails, considering changes in regulatory requirements, new risk scenarios, and feedback from system monitoring.
- Ensure that human oversight is involved when adjustments to guardrails are necessary, particularly when dealing with high-impact or high-risk decisions.
b. Document Adjustments and Interventions:
- Maintain a log of all guardrail adjustments and system interventions, including the rationale for changes and the outcome of each intervention.
- Conduct periodic audits to evaluate the effectiveness of guardrails and identify opportunities for improvement.

1. Verify that the Cloud Services Provider (CSP) has defined processes, procedures, and technical measures to apply guardrails to the AI system. Ensure that the processes are documented in detail, covering scope, objectives, roles and responsibilities.

2. Examine whether the above-mentioned processes, procedures, and technical measures are compliant with relevant regulatory requirements and industry best practices.

3. Confirm that the above-mentioned processes, procedures, and technical measures are concretely and appropriately implemented.

4. Inspect whether the above-mentioned processes, procedures, and technical measures are monitored against sets of efficacy and efficiency metrics / indicators.

5. Inspect whether the above-mentioned processes, procedures, and technical measures are periodically reviewed and updated by responsible parties.