Technical summary
SmartVerify: Execution Governance for AI
SmartVerify controls and proves AI actions on enterprise data. Interceptors at the tool server and data store join the person, agent, tool call, executed query and returned data in one signed entry per action, held in the customer's account. The Platform coverage grid is the source of truth for availability.
SmartVerify complements prompt guardrails, which evaluate requests, identity and IAM tools, which govern access, and DSPM tools, which map data posture. It adds a signed PostgreSQL execution record at a different enforcement point. See the security layer comparison for the evidence, enforcement point, timing, and availability of each layer.
Architecture and specifications
| Deployment model | Two interceptors join observations at the tool server and data store into one signed record per action. Single-interceptor deployment remains supported. No application-code changes. Start in record-only mode and add enforcement in stages. |
|---|---|
| Latency | Measured enforcement overhead: under 50 milliseconds at p90.Measurement scopePostgreSQL, two interceptors, without redaction. Results depend on deployment and enabled controls. |
| Traffic inspected | See the source-by-source availability labels in the coverage section on the Platform page. |
| Enforcement actions | Record and Refuse are available on PostgreSQL (including Amazon RDS and Aurora), MySQL, Amazon S3, Amazon DynamoDB, MCP and JSON-RPC. Reshape is available on PostgreSQL and MySQL. Reason across sessions, behavior-based cumulative limits, and Reshape for S3 and DynamoDB are on the roadmap. |
| Audit record | One signed entry per action joins the person, agent, tool call, query as executed and returned data. Stored in the customer's account; SIEM export by webhook. |
| Identity | Agents and tool servers receive short-lived, verified workload identities. Neither holds a database secret. |
| Limits | Customer policy defines records per query and tool-call limits per agent and person. Cumulative limits based on SmartVerify behavior analysis are on the roadmap. |
| Behavioral baselining | On the roadmap. The future Reason phase is intended to compare requests against the agent's established pattern of behavior. |
| Deletion propagation | Not available today. SmartVerify does not currently propagate deletion requests across RAG pipelines, vector stores, or caches. |
| Supported environments | AWS, Azure, Google Cloud, on premises and air-gapped customer environments. |
| Claude / Anthropic support | See MCP and JSON-RPC in the Platform coverage grid. |
How a request flows
- An AI agent issues a PostgreSQL statement toward a database.
- The SmartVerify inline edge worker observes the statement as it executes.
- The database returns the result over the observed connection.
- SmartVerify records the rows returned, columns read, session and agent identity.
- The per-session record is signed at the moment of observation.
Model Context Protocol (MCP) security
MCP and JSON-RPC observation and tool-call refusal are available now. See the canonical MCP tool-call refusal definition. SmartVerify records the tool invoked, the arguments given, and the call forwarded onward when the tool server is reachable through the customer network path. Read the MCP and JSON-RPC implementation guide.
Core terminology
- Prompt-to-execution drift
- The difference between what an AI agent was asked to do and the database statement that actually ran.
- MCP tool-call refusal
- A request-path control that stops an MCP tool call before the tool server acts on it and returns a typed error.
Implementation guides
- PostgreSQL execution evidence: architecture, record contents, role-based write refusal, and limitations.
- MCP and JSON-RPC security: network position, tool-call records, refusal flow, and limitations.
Regulatory mapping
| NIST SP 800-53 | Alongside NIST AI RMF, serves as a mapping hub from per-action evidence to AI and security frameworks and legal requirements. Mapping supports audits; it is not a certification or legal opinion. |
|---|---|
| NIST AI RMF | Per-query execution evidence maps onto relevant NIST AI RMF monitoring and logging sub-categories. |
| EU AI Act | Signed data-path records map onto logging and audit trail considerations for observed PostgreSQL activity. |
| Colorado AI Act (ADAI) | Execution evidence maps onto monitoring considerations for AI data access patterns. |
| California DROP | The regulation addresses deletion requirements across applicable systems. SmartVerify evidence maps onto the data-path record relevant to this consideration. |
| SOC 2 / HIPAA / PCI DSS / CMMC / GDPR | Signed, per-query execution evidence maps onto relevant controls in these frameworks. |
Framework mapping identifies controls and evidence supported by the SmartVerify record. It does not provide a certification or legal opinion.
Phases
| RecordAvailable now | Join the person, agent, tool call, query as executed, and returned data in one signed entry per action, stored in the customer's account. Retention is configurable, and longer periods with an immutable export are available for estates with a mandated retention period. |
|---|---|
| RefuseAvailable now | Refuse a query or tool call that breaks customer policy before it reaches the data or tool server. The agent receives an error its framework already handles. |
| ReshapeAvailable now | Apply policy to returned data so the agent receives only the rows and fields the request permits. Available on PostgreSQL and MySQL; S3 and DynamoDB reshaping are on the roadmap. |
| LimitAvailable now | Cap records per query and tool calls per agent and per person. Customers define the policy limits and counting windows. Behavior-based cumulative limits are on the roadmap. |
Record and Refuse are available on PostgreSQL (including Amazon RDS and Aurora), MySQL, Amazon S3, Amazon DynamoDB, MCP and JSON-RPC. Reshape is available on PostgreSQL and MySQL. Reason across sessions, behavior-based cumulative limits, and Reshape for S3 and DynamoDB are on the roadmap.
Company
SmartVerify, Inc. is headquartered in Bellevue, Washington. Built by the team that built, operated and scaled security and compliance services at AWS. Contact: smartverify.ai/contact. Whitepapers and research: smartverify.ai/resources.