Inputs that change intent
Consider whether manipulated requests or instructions in reference documents could redirect the service from its intended purpose.
AI Model & Service Security Testing
my-AI RedTeam
Examine the security boundaries of your AI service.
A solution for examining security risks that AI models and services may expose under varied inputs and usage conditions.
View full-size imageThree perspectives on AI security
- Inputs & instructions
- Data & access
- Policy & responses
Review examples · scope agreed for your environment.
Risks that ordinary answers may not reveal
Connected data and access
Consider the information that may appear in responses, alongside access to connected data and services.
Policy and actual responses
Review whether responses remain consistent with organizational policies and safety criteria in unexpected usage situations.
Review perspectives and use of findings
These are representative AI red-teaming perspectives, not a confirmed feature list. Applicable tests and supported scope are agreed for the target environment.
Instruction manipulation
Consider whether instructions in user inputs or reference documents could redirect the service or induce unauthorized responses and actions. Prompt injection is one example of this risk.
Sensitive data and access
Consider potential disclosure of personal or internal information beyond a user’s permissions. Connected data and application access controls matter alongside model responses.
Policy and response consistency
Examine whether responses to prohibited or reformulated requests align with organizational operating policies and safety criteria, and identify where protective measures are needed.
Prioritization and retesting
Findings can inform impact assessment and remediation priorities. Rechecking the same conditions after changes to policies, permissions or input and output handling helps assess improvement.
How we agree the review scope
Understand the target and purpose
Review the intended use and environment of the AI model or service you want to examine.
Agree perspectives and support scope
Discuss relevant review items, access conditions and the scope that can be supported.
Plan how to use the findings
Agree how to document risk conditions and improvement directions, and how to use them for follow-up review.
What we review before application
- Target AI model or service and its intended use
- Access conditions and authorization to review
- Review items and support scope
- Use of findings and follow-up review approach
Start with your AI review target and scope.
Discuss applicable review items and support scope for your environment and intended use.