Security intake for custom AI agent platforms.

Business context, users, tenants, and highest-risk agent behavior.

• Product name and short description
• Target users and customer types
• Current stage: prototype, beta, production, or customer-facing production
• Single-tenant or multi-tenant model
• Sensitive or regulated data the system may handle
• What the agent can see, decide, and change
• Worst realistic failure scenario

How requests, prompts, files, tools, model responses, and data move through the system.

• Frontend, Vercel API/serverless, backend services, database, storage, queues, and webhooks
• Where the custom agent loop runs
• How long-running jobs are handled
• Which environments exist: development, preview, staging, production
• Third-party services and integrations

Least-privilege access needed to validate the implementation.

• Staging/test URL
• Test user account
• Test admin/support account if applicable
• Test tenant with sample data
• Source repository or source export
• Vercel project settings export or screenshots
• Database schema, migrations, and policy exports
• Logging/monitoring samples or dashboard access

Identity, roles, tenant mapping, session handling, and enforcement points.

• Auth provider and flow
• Role and permission model
• Where authorization is enforced
• Tenant/user/account mapping
• Admin/support access and impersonation model
• Session expiration, revocation, MFA, invites, and password resets

Schema, row-level controls, backups, and cross-tenant protection.

• Database provider and version
• Tenant isolation design
• Row-level security or equivalent policy model
• Backend validation rules
• Backup/restore process
• Data export and deletion process
• Production data access list

Prompts, custom orchestration, tools, memory, retrieval, and action guardrails.

• System prompts and instruction layers
• Agent orchestration code
• Tool/function definitions
• Tool allowlists by tenant/user
• Memory and retrieval configuration
• Human approval gates for risky actions
• Model provider, models, and data retention settings
• Behavior tests or evaluations

Expected behavior under hostile prompts, malicious files, and unsafe tool requests.

• Ignore previous instructions / reveal system prompt
• Malicious uploaded document
• Malicious webpage summarized by the agent
• Cross-tenant ID manipulation
• Attacker-controlled webhook request
• Secrets/logs/schema disclosure request
• Destructive action request
• Cost spike or infinite-loop attempt

Route inventory, schemas, auth requirements, signature checks, replay protection, and rate limits.

• API route list
• Auth requirements per route
• Request and response schemas
• Webhook endpoint list
• Signature verification and replay protection
• Idempotency handling
• Rate limits by route, user, tenant, and IP

Upload limits, private storage, signed URLs, file scanning, embeddings, and deletion.

• File storage provider
• Allowed file types and size limits
• Private access model
• Signed URL expiration
• File scanning/validation
• Whether files are sent to models or vectorized
• Deletion of files and embeddings

Environment variables, key rotation, token scope, git history, and blast radius.

• Redacted environment variable inventory
• Secret storage locations
• Key rotation process
• Third-party token inventory
• Production access list
• Git secret scanning results
• Blast radius of model, database, and webhook key leakage

Request logs, agent/tool-call logs, redaction, alerts, audit trails, and kill switches.

• Sample request/auth/agent/tool-call logs
• Sensitive data redaction approach
• Who can view logs
• Alert rules for auth failures, errors, costs, data exports, and webhook failures
• Admin audit logs
• Incident response process
• Kill switch for compromised agents, tenants, integrations, or tools

Policies, subprocessors, model-provider data use, retention, consent, and regulated workloads.

• Privacy policy and terms
• Data retention policy
• Subprocessor list
• Data sent to model providers
• Model training/retention settings
• Deletion/export request support
• HIPAA, COPPA, FERPA, PCI, GLBA, SOC 2, GDPR, CCPA obligations