Security
Know every AI endpoint you run, and what happens when someone attacks it.
Nexus Cerberus registers AI endpoints, red-teams them against the OWASP LLM Top 10 mapped to MITRE ATLAS, and produces an AI-BOM, so a model reaches production with the same scrutiny as any other application.
- of every known AI endpoint
- Registryof every known AI endpoint
- Top 10 mapped to MITRE ATLAS
- OWASP LLMTop 10 mapped to MITRE ATLAS
- model supply-chain inventory
- AI-BOMmodel supply-chain inventory
- reporting on findings
- Signedreporting on findings
The operational problem
Models reach production faster than security teams learn they exist.
A product or data science team stands up a model endpoint to ship a feature. Security finds out when it appears in a support ticket, not through a review process. What breaks is the assumption that the AppSec pipeline already covers AI, when most of it was built for conventional application code.
The audit gap is a missing inventory: nobody can produce a current list of what models are running, what data they were trained or grounded on, or whether they have ever been tested against a known attack category.
What a solution must do
Treat AI endpoints as assets that need a register and a test cycle.
- Discovery
- Find AI endpoints across the estate, including ones not provisioned through a central process.
- Testing
- Attack them with adversarial prompts covering known LLM vulnerability categories, not just conventional web application checks.
- Supply chain
- Record model provenance, version and dependencies in a form that can be reviewed like a software bill of materials.
- Reporting
- Produce findings that a security team can act on and a leadership team can be shown.
How Nexus delivers it
Nexus Cerberus: registry, red-teaming, AI-BOM.
AI endpoint registry
Cerberus maintains a register of known AI endpoints across the estate, including model version and owning team, so security work starts from a real inventory rather than a survey.
AI red-teaming
Adversarial testing runs structured prompt-based attacks against a registered endpoint to surface prompt injection, data leakage, jailbreak and other model-specific weaknesses.
OWASP LLM Top 10 mapped to MITRE ATLAS
Findings are categorised against the OWASP LLM Top 10 and mapped to MITRE ATLAS tactics, giving security teams a shared vocabulary with their existing adversarial ML and threat intelligence work.
AI-BOM
An AI bill of materials records the model, its version, its training or grounding sources and its dependencies, supporting supply-chain review in the same way a software bill of materials does for code.
Signed reporting
Findings are compiled into a signed report suitable for a security review board or an audit request.
Where Cerberus sits
Cerberus assesses and reports. It registers endpoints, tests them adversarially, records the model supply chain and issues signed findings. It does not sit inline in the request path, so it never intercepts or blocks live model traffic.
Worked example
A support model deployed outside the formal review process.
- 01
Discover
Cerberus's endpoint registry finds a customer-support model endpoint deployed by a product team without a formal security review.
- 02
Inventory
The endpoint, its model version and its dependencies are recorded in an AI-BOM.
- 03
Red-team
Cerberus runs adversarial prompts against the endpoint, covering prompt injection, data leakage and jailbreak categories from the OWASP LLM Top 10.
- 04
Map
Findings are mapped to MITRE ATLAS tactics, so the security team can reason about them alongside other adversarial ML findings.
- 05
Report
A signed report is produced summarising exposures, severity and the affected endpoint.
- 06
Remediate
The product team addresses the highest-severity findings before the endpoint is approved for production traffic.
Evidence and output
What a Cerberus assessment produces.
AI-BOM summary
Illustrative entries
- support-assistant-v2OWASP LLM01 prompt injectionHigh
- support-assistant-v2OWASP LLM06 sensitive info disclosureMedium
- invoice-classifier-v1AI-BOM: 2 dependencies, 1 fine-tune sourceRecorded
- invoice-classifier-v1Last red-team assessmentPassed
Illustrative output. Actual reports include full finding detail and remediation guidance.
Products and platform capabilities
Cerberus, with platform sovereign AI as the thing being tested.
What Cerberus does
- AI endpoint registry
- AI red-teaming against the OWASP LLM Top 10
- Mapping to MITRE ATLAS
- AI-BOM and model supply-chain scanning
- Signed reporting
Where the boundary sits
- Inline interception or blocking of live model traffic in the request path
Who owns it
Security owns the programme, model owners own remediation.
Application security or AI security lead
Runs the registry and red-teaming programme, and owns the reporting cadence.
Model or product owner
Registers new endpoints, receives findings and is accountable for remediation before production release.
Risk and compliance
Uses the AI-BOM and reports to answer supply-chain and model-risk questions from auditors and regulators.
Architecture session
Find out what AI endpoints are already running.
We will help you stand up an endpoint registry and a red-teaming cadence, aligned with the OWASP LLM Top 10 and MITRE ATLAS, before the next model reaches production unreviewed.