There should be an AI Usage Policy in place that governs the use of AI tools in company workflows.

Control Type: Administrative

Control Function: Preventive

The 2020s have seen a rapid increase in the functionality of AI tools. Organizations of all sizes have seized the opportunity and integrated Ai into their environment to enhance various workflows. However AI can spin out of control and introduce new risks if not properly managed. Vulnerabilities in AI models continue to be uncovered and exploited. Since it is such a new technology, even seasoned cybersecurity professionals are behind on properly comprehending and addressing AI threats. Another issue with AI is how it can proliferate in organizations, even ones that don't directly sanction the use of AI for workflows. It is easy to create an account on ChatGPT or Grok. Employees can then input sensitive company data into LLMs without the organization's knowledge. This has become a major security issue in many organizations, referred to as Shadow AI. It is highly recommended that every organization establish at least a basic AI Usage Policy, even if AI use is not directly sanctioned. An AI Usage Policy should be regularly reviewed and updated as the AI industry and AI threat surface evolves.

An AI Usage Policy should contain:

  • Purpose & Scope: summarizes the relevance of AI in the environment and clarifies which departments and roles are subject to AI governance.
  • Approved AI Tools/Products: if the organization has professional subscriptions to specific AI platforms, or at the very least approves of the use of specific tools, these should be specified. The policy should clearly convey that all other AI tools are off limits.
  • Acceptable Use Scenarios: the organization needs to be able to articulate which workflows and tools AI can be used for. These may include data analysis, document proofreading, troubleshooting assistance, and any other activities the organization decides AI usage is acceptable for.
  • Prohibited Use Scenarios: in general, AI usage should be prohibited for all but the previously specified acceptable scenarios. However an organization may want to highlight specific scenarios that are off limits for AI assistance. These may include use in decision making processes, use with financial data, use with any data in the confidential or restricted security categories, and any activity that breaks laws or regulations.
  • Data Classification Rules: the policy should define which data from which classification levels is allowed to be entered into AI tools like Large Language Models (LLMs). Generally, public data can be used freely while Internal data can only be used by the approved AI tools. Confidential and Restricted data, especially PII and PHI should always be off limits. However there may be certain company-specific exceptions in the Confidential category that senior management feels should be allowed for certain requirements. In such cases, exceptions should be explicitly defined.
  • Security Requirements & Controls: Information security teams often need to get creative with how they lock down AI usage. One of the most effective controls is Data Loss Prevention (DLP) which can be used to block attempted leakage of sensitive data into AI tools. Organizations could also compile a hosts file or DNS filter that blocks access to all AI host websites other than the approved ones. This would prevent employees from accessing unauthorized platforms under the radar. Models also produce compliance logs that can be ingested into SIEM appliances to detect data leaks and unauthorized activity. Implementing these specific controls is out of scope of this piece of documentation, but any such controls that are implemented will need to be defined in the policy.
  • Human Review: AI security threats are an interesting niche because they are often enabled via human vectors rather than strictly technical ones. Overreliance on AI can end up being one of the most catastrophic threats to an organization, as poor decisions influenced by AI can infect all aspects on the organization down to the core. To mitigate this, the policy should mandate human reviews, preferably by multiple people, for decisions that are heavily influenced by AI.
  • AI Generated Content Governance: one of the most popular uses of AI, specifically Generative AI, is to generate visual marketing materials for organizations. While this can be a huge enhancement to an organization's public appearance, there should be a set of standards in place to keep the use of Generative AI in order. There should be a process for verifying factual accuracy, checking for bias, and preventing the insertion of harmful or malicious content.
  • Roles & Responsibilities: Lastly, the policy should specify who is responsible for managing and validating the proper usage of AI throughout the organization. Department level management can serve as the trusted authorities on proper AI use for their respective departments. Those managers are required to act in compliance with the standards outlined in the AI Usage Policy, and escalate larger issues or concerns to IT/cybersecurity managers, who in turn may need to escalate issues to higher levels of management.