Skip to content

Latest commit

 

History

History
256 lines (182 loc) · 11 KB

File metadata and controls

256 lines (182 loc) · 11 KB

Agent Data Handling Policy (ADHP) Specification

Version: 0.1.0 (Draft) Status: RFC — Request for Comments Author: AgentLedger Project Date: February 2026

1. Purpose

In a multi-agent ecosystem, when an orchestrator sends data to an agent for processing, the sender needs to know: what will happen to my data?

Traditional security frameworks (ISO 27001, GDPR) focus on who can access data. The Agent Data Handling Policy (ADHP) addresses a different question specific to the agentic world: what does the agent do with the data during and after processing?

This specification defines a standardized way for AI agents to declare their data handling practices, enabling orchestrators to make informed routing decisions based on the sensitivity of the data being processed.


2. Problem Statement

When Agent A sends a financial document to Agent B for analysis:

  • Will Agent B use that document to train or improve its model?
  • Will Agent B store the document after returning the analysis?
  • Will Agent B log the contents of the document?
  • Will Agent B forward the document to a sub-agent (Agent C)?
  • Will Agent B's response contain the original sensitive data?
  • Will Agent B share the data with third parties?
  • Where physically is Agent B processing the data?

Today, there is no standard way for agents to declare this information, and no standard way for orchestrators to filter agents based on these criteria.


3. Data Handling Levels

ADHP defines five levels of data handling, from most permissive to most restrictive.

3.1 Summary Table

Level Label Training Persistence Logging Delegation Output Third-Party Sharing
0 open Yes Unlimited Full Unrestricted May contain source data Allowed
1 standard No Session-based Metadata only Same level required May contain derived data With consent
2 sensitive No Request only Metadata only Same level+, must be declared Sanitized — no source data Anonymized only
3 strict No Request only No content logging Same level+ only Sanitized + reviewed Not allowed
4 zero-trace No None (streaming only) None No delegation Sanitized, no data leaves agent Not allowed

3.2 Detailed Definitions

Level 0 — Open

The agent makes no guarantees about data handling. Data may be used for any purpose including model training, indefinite storage, and redistribution. This is the default assumption when an agent does not declare a policy.

Use case: Public data processing, open-source analysis, non-sensitive content generation.

Level 1 — Standard

The agent will not use the data for model training. Data is retained for the duration of the session and then deleted. Only metadata (timestamps, request IDs, token counts) is logged — not the content itself. The agent may delegate to sub-agents, but only those that also operate at Standard level or above.

Use case: General business operations, non-regulated internal data, standard API calls.

Level 2 — Sensitive

Same protections as Standard, with additional constraints: data is retained only for the duration of a single request (not the full session). If the agent delegates to sub-agents, this must be explicitly declared in the agent's manifest. The agent's output is sanitized — it will not contain verbatim source data.

Use case: Personal data processing, financial analysis, HR data, customer records.

Level 3 — Strict

The agent provides strong confidentiality guarantees. No content logging of any kind. Data is retained only during request processing. Delegation is only permitted to agents at the same level or higher. Output is sanitized and reviewed (automated or manual) to prevent data leakage. No third-party sharing of any kind.

Use case: Legal documents, trade secrets, medical records, classified business strategies.

Level 4 — Zero-Trace

The highest level of confidentiality. The agent processes data in memory only — no disk writes at any point. No logging whatsoever. No delegation to any other agent. Output is sanitized and constrained so that no source data leaves the agent boundary. This level is designed for data that must leave absolutely no trace.

Use case: National security, pre-announcement M&A data, whistleblower submissions, highly sensitive IP.


4. Data Handling Properties

Beyond the overall level, each agent declares specific properties in its manifest:

4.1 Core Properties

Property Type Description
level enum One of: open, standard, sensitive, strict, zero-trace
training_opt_out boolean Whether the agent commits to NOT using data for model training
max_retention enum How long data is kept: none, request, session, 24h, 30d, unlimited
content_logging boolean Whether the content of requests/responses appears in logs
delegation_policy enum Sub-agent policy: none, same_or_higher, unrestricted
output_sanitization boolean Whether outputs are scrubbed of source data
jurisdiction list[string] Where data is processed (ISO 3166-1 country codes)
certification string (nullable) Future: ID of a verification/audit certificate

4.2 Third-Party Sharing Properties

Property Type Description
third_party_sharing.enabled boolean Whether data is shared with any third party
third_party_sharing.purpose list[enum] Why data is shared: analytics, advertising, improvement, subprocessing, resale
third_party_sharing.sanitized boolean Whether shared data is anonymized/stripped of PII
third_party_sharing.parties_disclosed boolean Whether the list of third parties is publicly available
third_party_sharing.opt_out_available boolean Whether the data sender can opt out of sharing

4.3 Example Manifest

{
  "agent_name": "FinanceAnalyzer Pro",
  "data_handling": {
    "level": "strict",
    "training_opt_out": true,
    "max_retention": "request",
    "content_logging": false,
    "delegation_policy": "same_or_higher",
    "output_sanitization": true,
    "jurisdiction": ["DE", "FR"],
    "certification": null,
    "third_party_sharing": {
      "enabled": false,
      "purpose": [],
      "sanitized": true,
      "parties_disclosed": true,
      "opt_out_available": true
    }
  }
}

5. Delegation Cascading Rule

Critical principle: When Agent A delegates work to Agent B, the data handling level must be maintained or strengthened — never weakened.

Agent A Level Agent B Minimum Level
open open (any)
standard standard
sensitive sensitive
strict strict
zero-trace No delegation allowed

The registry MUST enforce this rule: when an orchestrator queries for agents to handle sensitive data, it must only return agents whose full delegation chain maintains the required level.

This is a key value proposition of the registry — it acts as a trust broker that verifies the entire chain of data handling.


6. Third-Party Sharing Matrix

Level Sharing Allowed Conditions
open Yes No restrictions
standard Yes Only with sender's consent
sensitive Limited Only anonymized/sanitized data
strict No Third-party sharing prohibited
zero-trace No Third-party sharing prohibited, no data leaves agent

7. Verification Roadmap

Phase 1 — Self-Declaration (MVP)

Agents declare their own data handling level. The registry stores this declaration. Orchestrators filter based on declared levels. Trust is based on the agent operator's reputation.

Phase 2 — Verified Badge

Agent operators can request verification. This involves:

  • KYC (Know Your Customer) for the operating organization
  • Technical audit of the agent's data handling practices
  • Periodic re-verification
  • A "Verified" badge displayed in the registry

This is a monetization path — organizations pay for verification, which increases their visibility and trust score in the registry.

Phase 3 — Automated Auditing

Trusted "Auditor" agents periodically test registered agents by:

  • Sending test data with tracking markers
  • Verifying the data is handled according to the declared policy
  • Checking that delegation chains maintain confidentiality levels
  • Reporting violations automatically

Phase 4 — Cryptographic Verification

Technical enforcement through:

  • Encrypted data envelopes that enforce retention policies
  • Cryptographic proofs of deletion
  • Secure enclaves for zero-trace processing
  • Blockchain-based audit trails for verification history

8. API Integration

Registering an Agent with Data Handling Policy

POST /agents
{
  "name": "FinanceAnalyzer Pro",
  "description": "Advanced financial document analysis",
  "capabilities": ["financial-analysis", "document-parsing"],
  "endpoint": "https://api.finanalyzer.example.com",
  "protocol": "REST",
  "data_handling": {
    "level": "strict",
    "training_opt_out": true,
    "max_retention": "request",
    ...
  }
}

Querying Agents by Data Handling Requirements

GET /agents?min_data_handling_level=sensitive&jurisdiction=EU&third_party_sharing=false

Returns only agents that meet or exceed the requested confidentiality requirements.


9. Default Behavior

  • If an agent does not declare a data_handling policy, the registry defaults to level: "open"assume the worst case if not declared.
  • This incentivizes agents to explicitly declare their policy, as undeclared agents will be filtered out of any confidentiality-sensitive queries.

10. Relationship to Other Standards

Standard Focus ADHP Complement
ISO 27001 Who can access data within an organization What happens to data after an agent processes it
GDPR Legal framework for personal data in the EU Technical declaration of compliance-relevant practices
SOC 2 Organizational security controls Agent-level data handling transparency
OAuth/OIDC Who is authorized to call the agent Authorization is separate — handled by external auth systems
MCP Protocol for agent-tool communication ADHP adds data handling metadata to the agent discovery layer

ADHP is designed to complement, not replace, these standards. An agent operating at "strict" level within the EU would still need to comply with GDPR independently — ADHP simply makes the agent's practices transparent and machine-queryable.


Contributing

This specification is in draft. We welcome feedback on:

  • Are the five levels sufficient? Too many? Too few?
  • Are there data handling concerns we've missed?
  • How should verification work in practice?
  • How does this interact with emerging AI regulation?

Please open a Discussion or submit an RFC in the docs/ folder.