AI Responsible Use & AI Security Standard

This standard sets the minimum requirements Raeburn Consulting applies when artificial intelligence is used in client work, internal operations, software products, discovery tooling, automation, analysis or content generation. It is intended to make our AI use understandable, reviewable and proportionate to risk.

1. Client data use

Client data may be used only for the agreed service purpose, by authorised systems and people, and subject to applicable contractual, privacy and security requirements. We minimise the data supplied to AI systems and do not intentionally provide credentials, secrets or data that is unnecessary for the task.

2. Model training

Client content is not treated as a general-purpose Raeburn training corpus. We do not intentionally use client confidential information to train or fine-tune Raeburn models for unrelated customers without an explicit, documented lawful basis and appropriate client agreement. Where a third-party AI provider is used, its data-use and training terms form part of provider selection and assurance.

3. Prompt and data retention

Prompts, inputs, outputs and supporting data are retained only where there is a defined operational, security, contractual or legal purpose. Retention should be minimised and aligned with the relevant service or client record. Provider retention settings are considered during provider selection and configuration.

4. Subprocessors and AI providers

AI model providers, hosting platforms and other processors are treated as suppliers. Selection considers security, privacy, contractual terms, data location, retention, model-data use, incident handling and the sensitivity of the intended workload. Material subprocessors are handled through our supplier-assurance and privacy processes.

5. Human review and accountability

AI output is not assumed to be correct merely because it is fluent or plausible. Human review is required where an output could materially affect a client, an individual, a legal or financial position, security, safety, contractual commitments or other high-impact decisions. Accountability remains with the responsible human or service owner, not the model.

6. Hallucination and reliability risk

AI systems can fabricate facts, citations, calculations or reasoning. Material claims should therefore be checked against authoritative evidence appropriate to the task. Where reliable verification is not possible, uncertainty must be made clear rather than presented as fact.

7. Sensitive and regulated data

Sensitive personal data, special-category data, authentication material, payment information, privileged material and highly confidential client information require heightened controls. They must not be placed into an AI service merely for convenience. Use must be necessary, authorised and compatible with the approved provider, contract and privacy requirements.

8. Access control

Access to AI-enabled systems, prompts, customer context, model credentials and administrative functions follows least-privilege principles. Privileged functions must be separated from ordinary user content and should be authenticated, authorised, revocable and logged where technically appropriate.

9. Model and provider selection

Model selection is based on the intended use rather than capability alone. We consider security, privacy, provider terms, data handling, model suitability, reliability, safety features, operational resilience, cost and the ability to apply appropriate controls. Higher-risk workloads require stronger assurance.

10. Logging and auditability

AI-enabled services should retain sufficient security and operational evidence to investigate misuse, failures and incidents without unnecessarily logging confidential prompt content. Where possible, events are recorded using metadata, identifiers, hashes or other minimised evidence rather than duplicating sensitive source material.

11. AI security testing

AI-enabled applications are included in secure-development and vulnerability-management processes. Testing is risk-based and may include input validation, authorisation tests, prompt-injection testing, unsafe-output handling, dependency and code scanning, abuse-case testing and production verification where applicable.

12. Prohibited uses

Raeburn AI systems must not be intentionally used to facilitate unlawful activity, unauthorised access, credential theft, malware, deceptive impersonation, discriminatory decision-making without appropriate safeguards, covert surveillance, deliberate privacy violations, or actions designed to bypass security controls. AI must not be given autonomous authority for high-impact actions unless the relevant system has explicit safeguards, authorisation and oversight.

13. Copyright and intellectual property

AI does not remove normal copyright, licensing, confidentiality or intellectual-property obligations. Inputs must be used lawfully, and material outputs intended for external or commercial use should be reviewed for provenance, licensing, substantial reproduction and client-specific IP requirements where relevant.

14. Prompt injection

External documents, web content, emails, retrieved text and client-supplied material are treated as untrusted data, not trusted instructions. AI systems should separate privileged instructions from untrusted content and must not follow embedded requests to reveal secrets, alter security policy, expand permissions or perform unauthorised actions.

15. Data exfiltration prevention

Models must not be given unrestricted access to secrets or customer data simply because an AI workflow could use them. Tools and connectors should expose only the minimum data and actions required. Sensitive outputs, tool calls and cross-tenant access are subject to authorisation and validation controls appropriate to the service.

16. Model output review

Outputs used for client deliverables, decisions, code changes or consequential actions must be reviewed to a level proportionate to their impact. Structured outputs should be validated against expected schemas or business rules where feasible. Unsafe, malformed or unverifiable output must be rejected, corrected or escalated rather than silently accepted.

Governance, incidents and questions

Suspected AI security issues, prompt-injection vulnerabilities, data leakage or unsafe AI behaviour can be reported to security@theraeburngroup.com. Privacy questions can be directed to dpo@theraeburngroup.com. Our Security & Responsible Disclosure Policy applies to good-faith vulnerability reports.

This standard is reviewed as our services, providers, threat landscape and regulatory obligations evolve. It describes minimum governance requirements and does not replace client-specific contractual controls where stronger requirements apply.