Reference
Policies for OpenAI
See the OpenAI inventory reference for the objects these templates work on.
| Policy | Description | Severity | Category | Checks |
|---|---|---|---|---|
| External Entra ID users have access to OpenAI | Detects OpenAI user which are external users in the Entra ID tenant External users accessing AI services pose additional security and compliance risks, including potential data exposure to external parties. Organizations should carefully control and monitor external access to sensitive AI capabilities. | Medium | Security | OpenAI User |
| Admin API key is live (recently used) ⭐ | Recently used admin API keys are live, organization-wide credentials, raising the likelihood of serious impact Admin API keys carry organization-wide privileges. A recently used admin key is both live and privileged: if compromised it grants full administrative reach, and recent use confirms it is an active credential. This combination raises the likelihood that an access issue has serious impact. This raises the likelihood axis of the risk score, not the impact axis. | High | Security | OpenAI Admin API Key |
| OpenAI agent conversation with malicious request ⭐ | Detects a chat which might be used to extract sensitive information Chat conversations flagged as containing malicious requests may indicate prompt injection attacks, attempts to extract sensitive training data, or social engineering through AI systems. These incidents require immediate security review to assess potential data exposure, identify attack patterns, and implement protective measures. Administrators should investigate the conversation context, determine if sensitive information was disclosed, and consider restricting the user's access pending review. | High | Security | OpenAI ChatKit Thread |
| Archived OpenAI project with active API keys ⭐ | Detects archived projects that still have API keys which could be used for unauthorized access Archived OpenAI projects that still have active API keys represent a security risk. These keys can still be used to make API calls and incur costs even though the project is archived. Remove or revoke all API keys from archived projects to prevent unauthorized usage and reduce attack surface. | High | Security | OpenAI Project |
| Assistant is live and callable ⭐ | Provisioned assistants are API-callable AI surfaces, raising the likelihood that any weakness is exploited A provisioned assistant is a live AI surface: it can call the model and use tools, and is invocable by any holder of the project's API key. Any weakness in its instructions, tools or backing model is therefore exposed the moment the assistant exists. This raises the likelihood axis of the risk score, not the impact axis. | Medium | Operation | OpenAI Assistant |
| OpenAI Certificate expires soon ⭐ | Detects organization certificates which will expire soon and should be updated Certificates expiring within 90 days require immediate attention to prevent service disruptions. Certificate-based authentication is critical for API access, and failing to renew before expiration will cause authentication failures, blocking all operations dependent on that certificate. Plan certificate renewal well in advance to ensure continuity of AI service operations. | High | Security | OpenAI Organization Certificate |
| OpenAI Certificate is expired ⭐ | Detects organization certificates which are expired and should be updated Expired certificates cause immediate authentication failures and complete service outages for all API operations. This represents a critical security and operational issue that blocks access to OpenAI services and disrupts dependent applications. Replace expired certificates immediately and investigate why renewal processes failed to prevent recurrence. | High | Security | OpenAI Organization Certificate |
| ChatKit thread is active ⭐ | Active threads are live, ongoing conversations, raising the likelihood that a content or data issue occurs An active ChatKit thread is a live, ongoing conversation that continues to accept input and surface model output. Active threads are reachable and in use, raising the likelihood that a content or data-exposure issue actually occurs. Closed and locked threads do not match. This raises the likelihood axis of the risk score, not the impact axis. | Medium | Security | OpenAI ChatKit Thread |
| OpenAI ChatKit thread is locked | Detects chat threads with locked status indicating potential policy violations or security incidents Locked threads may indicate policy violations, content moderation triggers, or security incidents. These threads should be reviewed by administrators to understand why they were locked and whether further action is needed. | Medium | Security | OpenAI ChatKit Thread |
| Expired OpenAI invites should be cleaned up | Detects OpenAI invitations that have expired and should be removed Expired invitations clutter the admin view and may indicate stale onboarding processes. Cleaning up expired invites improves administrative hygiene and ensures the invite list reflects actual pending access requests. | Low | Declutter | OpenAI Invite |
| OpenAI File contains PII ⭐ | Detects uploaded files which contain Personally Identifiable Information for training Personally Identifiable Information (PII) uploaded to OpenAI services poses serious privacy and compliance risks under regulations like GDPR and CCPA. PII exposure in training data or file storage can lead to data breaches, legal penalties, and loss of customer trust. Files containing PII should be removed immediately to prevent unauthorized access and ensure regulatory compliance. | High | Security | OpenAI File |
| OpenAI File contains sensitive information ⭐ | Detects uploaded files which contain company, medical or financial data Files containing sensitive company, medical, or financial data uploaded to OpenAI services create significant confidentiality and compliance risks. Exposure of proprietary business information, healthcare records (HIPAA), or financial data (PCI-DSS, SOX) can result in competitive disadvantages, regulatory violations, and severe legal consequences. Immediate removal of such files is essential to protect organizational assets and maintain regulatory compliance. | High | Security | OpenAI File |
| File never expires ⭐ | Files with no expiry persist and stay reachable, raising the likelihood that sensitive content is exposed over time A file with no expiry persists indefinitely and stays retrievable for as long as the project lives. The longer sensitive content remains reachable, the greater the likelihood it is eventually exposed or exfiltrated. Files with an expiry date age out and do not match. This raises the likelihood axis of the risk score, not the impact axis. | Medium | Security | OpenAI File |
| OpenAI fine-tuning file contains PII ⭐ | Detects files used for fine-tuning that contain Personally Identifiable Information Files used for fine-tuning are incorporated into model training data. PII in fine-tuning datasets creates severe compliance risks under GDPR, CCPA, and HIPAA — the personal data may be memorized and later surfaced in model outputs. This represents the highest-risk data exposure scenario in the OpenAI platform. | Critical | Security | OpenAI File |
| OpenAI Organizations with too less owners ⭐ | Detects organizations that has less than 2 owners Having only one owner creates a single point of failure for organization management. If that owner becomes unavailable or leaves, administrative access and control may be lost, potentially disrupting critical operations. | High | Security | OpenAI Organization |
| OpenAI Organizations with too many owners ⭐ | Detects organizations that has more than 5 owners Excessive owner privileges expand the attack surface and dilute accountability. Each additional owner represents a potential security risk if their credentials are compromised. Limiting ownership ensures clear responsibility and reduces exposure. | High | Security | OpenAI Organization |
| OpenAI organization exceeds the 30 days costs limit | Detects OpenAI organizations with costs more than 5000 EUR in the last 30 days Organization-level cost monitoring is essential for budget governance. Uncontrolled AI spending across projects can lead to significant budget overruns. Best practice recommends per-organization spending caps aligned with the purchased tier. | High | Costs | OpenAI Organization |
| Stale pending OpenAI invites should be revoked | Detects OpenAI invitations that have been pending for more than 30 days Long-pending invitations may indicate abandoned onboarding or forgotten access grants. If the invite link is eventually found by an unauthorized party, it could grant unintended access. Stale pending invites should be revoked and re-issued if still needed. | Medium | Security | OpenAI Invite |
| Project API key is live (recently used) ⭐ | Recently used project API keys are live credentials, raising the likelihood that a weakness leads to an incident A recently used project API key is a live credential. If leaked it grants immediate access; idle keys are far less likely to be the source of an active incident. Recent use therefore raises the likelihood that any weakness — no rotation, broad scope — results in a real incident. This raises the likelihood axis of the risk score, not the impact axis. | Medium | Security | OpenAI Project API Key |
| OpenAI Project Certificate expires soon ⭐ | Detects project certificates which will expire within 90 days and should be updated Project-level certificates provide scoped authentication. Expiration causes project-specific API failures that may be harder to diagnose than organization-level outages. Plan certificate renewal well in advance to ensure continuity of AI service operations. | High | Security | OpenAI Project Certificate |
| OpenAI Project Certificate is expired ⭐ | Detects project certificates which are expired and should be updated Expired project certificates immediately block all API operations for that project. This is a critical operational issue requiring immediate remediation. Replace expired certificates and investigate why renewal processes failed. | Critical | Security | OpenAI Project Certificate |
| Project is live (active) ⭐ | Active, non-archived projects are reachable and in use, raising the likelihood that an issue manifests An active (non-archived) project is live: it holds usable API keys, files and ongoing usage. Live projects are reachable and in use, raising the likelihood that any cost or data issue actually manifests. Archived projects do not match. This raises the likelihood axis of the risk score, not the impact axis. | Medium | Operation | OpenAI Project |
| OpenAI projects exceed the 30 days costs limit | Detects OpenAI projects with costs more than 1000 EUR in the last 30 days High AI usage costs can indicate inefficient use, unnecessary model calls, or runaway processes. Monitoring cost thresholds helps prevent budget overruns and encourages optimization of AI resource consumption. | Medium | Costs | OpenAI Project |
| OpenAI projects exceed the 7 days costs limit | Detects OpenAI projects with costs more than 500 EUR in the last 7 days A sudden spike in 7-day costs may indicate runaway processes, inefficient model calls, or unauthorized API usage. Monitoring short-term cost thresholds enables early detection before monthly bills escalate significantly. | Medium | Costs | OpenAI Project |
| OpenAI Project without any members | Detects OpenAI projects that have no members assigned Projects without members are orphaned resources that may still have active API keys or service accounts. They increase governance complexity and may represent shadow IT or abandoned experiments. Orphaned projects should be reviewed and archived. | Medium | Declutter | OpenAI Project |
| OpenAI Project with too less owners ⭐ | Detects projects with less than 2 owners Projects with a single owner risk becoming inaccessible if that owner is unavailable. This creates operational risk and prevents proper continuity of AI development and deployment activities. | Medium | Security | OpenAI Project |
| OpenAI Project with too many owners ⭐ | Detects projects with more than 10 owners Too many project owners complicates access management and increases security risks. With numerous privileged users, tracking who made changes becomes difficult, and the risk of credential compromise or misuse escalates significantly. | High | Security | OpenAI Project |
| OpenAI User is deactivated in Entra ID | Detects OpenAI users that have been disabled in the tenant When users are disabled in Entra ID, their OpenAI access should be revoked immediately. Lingering access permissions for disabled accounts create security vulnerabilities and violate the principle of least privilege. | Medium | Security | OpenAI User |
| External (guest) user ⭐ | External guest accounts are a common attack vector, raising the likelihood of risky access External (guest) accounts operate outside corporate identity governance and are a common attack vector. Their presence raises the likelihood that an access weakness is actually exploited. Internal users simply do not match. This raises the likelihood axis of the risk score, not the impact axis. | High | Security | OpenAI User |
| Unused OpenAI admin API Key should be removed ⭐ | Detects OpenAI admin API Keys that have not been used within the last 90 days Admin API keys carry organization-wide privileges. Unused admin keys are high-value targets — if compromised, they grant full administrative access. Industry best practice recommends rotating or revoking admin keys every 60-90 days to reduce the attack surface. | High | Security | OpenAI Admin API Key |
| Unused OpenAI project API Key should be removed | Detects OpenAI project API Keys that have not been used within the last 90 days Unused API keys represent dormant security risks that can be exploited if compromised. Removing inactive keys reduces the attack surface and prevents potential unauthorized access through forgotten credentials. | Medium | Declutter | OpenAI Project API Key |
| Unused OpenAI Project should be archived | Detects OpenAI Projects without any completion within the last 90 days Inactive projects that accumulate without cleanup increase management overhead and complicate governance. Archiving unused projects simplifies the environment and reduces potential attack vectors from forgotten resources. | Medium | Declutter | OpenAI Project |