Managed model endpoints
AI services in your AWS, Azure and Google Cloud accounts, to be found from the same inventories SecurityVault already builds for cloud posture.
Coming soon
See your AI estate.
Focus on its security risks.
Build a shared view of AI services, model exposure and application risks. Plan targeted evaluations for prompt injection, sensitive-data leakage and unsafe integrations.
AI service inventory
AI endpoint
Model configuration
Access boundary
Investigation context
01 / Inventory AI
Define the services in scope.
- Service scope
- Deployment context
Discuss which AI assets and deployment configurations need review.
02 / Inspect exposure
Review the boundary around the service.
- Access policy
- Exposure question
Evaluate configuration and access context relevant to an AI service.
03 / Agree evaluation
Set a concrete review outcome.
- Review owner
- Success criteria
Define the evidence your team needs to evaluate AI posture coverage.
Find
AI discovery
Managed services and self-hosted models.
Catalogue
Model inventory
Owner, environment, data handled.
Test
LLM security tests
Injection, jailbreak, leakage.
Guard
Input and output filters
Will decode text before matching.
The category
What is AI-SPM?
AI-SPM inventories the AI services an organisation depends on, assesses their configuration, tests their behaviour and keeps the result current as models change. It covers familiar risks, such as an endpoint without authentication or an insecure model file format, and new ones, such as a model that can be talked into revealing its instructions or the personal data it was given.
Model surface
Identify the model and the application using it.
Security context
Review the data, access and configuration boundaries.
Remediation scope
Define which issue and asset need attention.
Designed workflow: review the offering’s scope and availability with our team.
Discovery
Where does AI risk enter your estate?
AI arrives through many doors at once, often without a security review. AI-SPM is designed to find each one and record it in a single model catalogue with its provider, version, environment, owner and the data it handles.
01
02
Self-hosted model servers
Model servers on your own GPU nodes, to be found from network scan results and assessed for exposure, authentication and logging.
03
Model supply chain
A model loaded from an untrusted source or an unsafe serialisation format is a supply chain issue, and will sit beside your dependency and image findings.
The design
What will AI-SPM test for?
Tests will run against a live endpoint. An endpoint that cannot be reached will be reported as an error, never as a pass.
| Test | What it will probe |
|---|---|
| Prompt injection | Instructions that override the system prompt, including indirect injection planted in documents, retrieval results or tool output |
| Jailbreaks | Multi-turn attempts to bypass the model's safety rules |
| Data extraction | Attempts to recover system instructions, context or training data |
| Personal-data leakage | Responses that disclose personal data the model was given |
| Content safety and hallucination | Harmful output and confidently wrong answers |
| Configuration | Endpoint exposure, authentication, logging, access control and insecure model formats |
From finding to fix
What happens when a model fails a test?
Ranked by exposure
An internet-reachable model endpoint without authentication will be ranked with the same exposure context as any other public workload, and a model that handles personal data and follows injected instructions will read as a data exposure.
Guarded at runtime
An input and output filter will decode encoded and look-alike text before matching, and check retrieved content for planted instructions.
Blocked, not only reported
AI policies are designed to run in the same policy engine as your fail-closed deployment gates, so an unauthenticated critical model can be stopped before release.
Owned and evidenced
Failed tests will reach an owner with a remediation deadline and feed continuous compliance evidence; accepting an AI risk will remain a human decision with a reason code.
FAQ
AI-SPM questions.
- What is the difference between AI-SPM and AI governance?
- AI-SPM is the security view: where models run, how they are exposed and whether they resist attack. AI governance is the oversight view: which uses are approved, under which policy, with what record of decisions. They share one model inventory. See AI governance.
- Which AI services will AI-SPM discover?
- It will discover managed AI services in your existing AWS, Azure and Google Cloud inventories, and self-hosted model servers in network scan results, and record each one in a model catalogue.
- What will the LLM security tests cover?
- Prompt injection, jailbreaks, data extraction, personal-data leakage, content safety and hallucination, plus indirect prompt injection planted in documents, retrieval results or tool output across multi-turn conversations. An unreachable endpoint will be reported as an error, never as a pass.
- Will AI-SPM protect the data used to train models?
- It is designed to work alongside data security posture management, which will scan training datasets for sensitive data and check their declared provenance before a training run.
- When will AI-SPM be available?
- AI-SPM is in development. Ask our team how your AI deployment and model inventory requirements align with the planned release scope.
Related offerings: AI governance · data security posture management. Available today: cloud security posture · API security. Built for security engineering and platform engineering teams.
Discuss your requirements
Tell us where AI runs in your estate.
Map your AI models, training-data concerns and deployment risks to the planned evaluation scope.