Accounts and access
Secure sessions, email verification, optional TOTP MFA, five workspace roles and tenant-scoped authorisation.
Trust Centre
Know exactly what is checked, stored and proven.
Security claims should be inspectable. This page separates product controls, evidence boundaries, operational checks and work that still requires independent assurance.
The evidence chain
Every deployment decision should be traceable from claim to retest.
AgentRiskLayer does not treat a questionnaire answer as technical proof. Each evidence class remains visible and separately labelled.
- 1Declared controls
What the customer says exists.
- 2Observed controls
What code, configuration or connected systems show.
- 3Findings
Where a supported exposure or control weakness remains.
- 4Attack evidence
What an authorised controlled test reproduces.
- 5Runtime evidence
What a live policy allowed, denied or paused.
- 6Human approval
Who approved the exact high-impact action.
- 7Remediation
Who owns the fix and what changed.
- 8Retest
Whether the same risk remains controlled.
- 9Deployment decision
Proceed, hold or do not deploy—with limitations.
Security and privacy controls
Controls customers can verify.
These are product controls, not claims of certification. Exact implementation and operational status should be checked against the current release and readiness evidence.
Fail-closed policy decisions
Versioned policies screen prompts, outputs and tool calls before an application acts.
Exact-action human control
High-impact approval is server-issued, expiring, bound to the complete action and consumed once.
Bounded hosted evidence
Hosted Guard evidence records decisions, rules and digests rather than raw prompts, outputs or tool arguments.
Integrity and traceability
Reports, Inspector bundles and controlled test artifacts include hashes or signatures where appropriate.
Readiness and recovery
Production readiness, health, protected metrics, database backup controls and documented restoration procedures.
Data boundary
What each product surface handles.
Different workflows need different data. The boundaries below prevent one broad claim from hiding important differences.
SurfaceWhat is processedWhat is retained
Risk checkYour selected answers and optional evidence labelsAssessment answers, findings and result metadata
Local InspectorAuthorised local files and configurationRedacted evidence facts and integrity metadata—not source contents or matched secret values
Controlled Red TeamAuthorised synthetic or staging interactionsBounded outcomes, rule references and redacted evidence
Hosted GuardPrompt, output or proposed tool call for immediate evaluationDecision, policy identity, rule IDs and digests—not raw prompt, output or arguments
Explicit limitations
What AgentRiskLayer does not claim.
- It is not an accredited certification.
- It does not guarantee that an AI system is risk-free.
- A questionnaire result is not a penetration test.
- Local evidence generated on a customer-controlled machine does not independently prove completeness.
- Synthetic regression results are not customer outcomes or independent performance evidence.
- Production security depends on the exact deployed release, configuration, integrations and operating practices.
Verification routes
Go directly to the evidence you need.
17,600-action benchmarkSee the frozen workload, observed outcome, artifact hashes and explicit limits →Security centreArchitecture, application controls and validation boundaries →Risk knowledge libraryBrowse each problem, control objective and evidence boundary →Assessment methodologyScoring, evidence confidence, attack paths and decisions →Standards mappingRecognised guidance linked to actual findings →System statusLive application and database readiness →Vulnerability disclosureHow to report a security issue privately →Controlled proof manifestReproducible synthetic evidence package →
See the boundary in practice
Watch the controlled demo