Naka: A Disclosure-Budget Guardrail for AI Agents
Naka is a disclosure-budget guardrail for AI agents that tracks cumulative sensitive-data exposure across tool calls. It combines Cedar authorization, PII detection, and a per-session disclosure ledger to prevent excessive data disclosure and unauthorized rehydration.
AI agents are becoming capable of reading customer records, calling enterprise tools, and combining information from multiple sources. This creates a privacy problem that traditional authorization does not fully address.
A conventional authorization system can answer:
“Is this agent allowed to make this tool call?”
But it does not necessarily answer:
“Should this principal be allowed to learn this field about this person, considering everything they have already learned during this session?”
We built Naka to address this gap.
What we built
Naka is a disclosure-budget guardrail for AI agents. It evaluates the agent’s tool call using authorization policies and then evaluates the information returned by the tool before it reaches the agent.
For every detected sensitive field, Naka considers:
- Who is requesting the information
- Which session the request belongs to
- Which person the information belongs to
- What sensitive fields have already been disclosed
- The remaining disclosure budget
- Whether the requested disclosure is permitted by policy
The system maintains this state in a per-subject, per-session disclosure ledger.
For example, a session with a disclosure budget of three could progress like this:
PAN → released → 1/3
GSTIN → released → 2/3
IFSC → released → 3/3
Voter ID → denied → budget exhausted
Each individual request may look reasonable in isolation. Naka considers the cumulative exposure.
Preventing rehydration-based exfiltration
We also addressed another important problem: replacing sensitive information with a placeholder should not create a new way to retrieve that information.
Naka treats sensitive-data rehydration as a separate authorization decision and restricts it to explicitly authorized destinations.
So a placeholder cannot simply be sent to an arbitrary external destination and redeemed there.
This creates a second security boundary around sensitive information.
AWS technologies we used
Naka combines AWS services with the open-source Cedar authorization framework.
Amazon Bedrock provides the AI/agent integration layer.
Amazon Comprehend is used as a PII detection tier for identifying sensitive information in tool results.
AWS Lambda provides the serverless execution layer for the guardrail and supporting functions.
Amazon DynamoDB stores the disclosure ledger and session state required for cumulative privacy decisions.
Amazon Cedar provides the policy-based authorization layer for tool calls, field disclosure, roles, and rehydration destinations.
Amazon CloudWatch provides operational monitoring and visibility during development and deployment.
We built the application and AWS integrations using Python and Boto3, with a web-based control plane for observing decisions, disclosure state, detector coverage, and policies.
What fought back
Building Naka also exposed some real-world constraints.
Some Amazon Bedrock and Amazon Comprehend capabilities were gated on our AWS account during development. This meant we could not depend on a fully autonomous live-agent loop or assume every detection tier would always be available.
Rather than hiding that limitation, we designed Naka so that detector availability is explicit.
If a detection tier does not run, the system records and displays that state instead of presenting a narrower scan as a complete one.
For the final demonstration, deterministic scenarios drive the actual guardrail components, including the Cedar authorization decisions, disclosure ledger, audit processing, and rehydration authorization.
We also had to carefully design the DynamoDB state model because this is not a normal access-control problem. The decision depends on the relationship between the principal, session, data subject, and previously disclosed fields.
Why Naka?
The core idea behind Naka is simple:
An AI agent’s privacy impact is cumulative.
When an agent makes several tool calls, each response can contribute another piece of information about the same person. Protecting each call independently is not always enough.
Naka brings that session history into the authorization decision.
Instead of asking only:
“Can the agent access this?”
Naka also asks:
“How much more should this principal be allowed to learn?”
That is the problem we set out to solve with Naka.
Built by Team Morolo.
#bharatBuilds #AWS #AWSHackathon #AIAgents #GenerativeAI #AmazonBedrock #AmazonCedar #AmazonDynamoDB #AmazonComprehend #AWSLambda #CloudSecurity #PrivacyEngineering
Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article