What Matters
Connecting an agent does not merely reveal the prompt. A wallet connection can expose addresses, balances and transaction history; an exchange API can expose orders, positions and account identifiers; a model provider may receive prompts and tool results; an orchestration platform may log every call; RPC, analytics and payment providers can observe metadata. Privacy depends on the entire data path, not whether one application says it is “self-hosted.”
DN Evidence Block
Reviewed: 19 November 2026. Primary evidence: NIST Privacy Framework; official OpenAI API data-control documentation; Anthropic API privacy and retention documentation; Model Context Protocol authorization guidance; public blockchain and exchange API behavior.
Boundary: provider defaults, retention periods and zero-data-retention eligibility change by product and contract. Verify the exact service, region, account tier and endpoint rather than assuming one policy covers an entire brand.
The Agent Privacy Map
| Data surface | What may leave your control | Who may observe it | Minimum control |
|---|---|---|---|
| Model prompt | Instructions, documents, wallet labels, strategy and exceptions | Model provider, gateway, observability platform and support staff | Redaction, approved fields and retention policy |
| Wallet connection | Addresses, balances, assets, signatures and transaction graph | dApp, RPC, indexer, chain observers and analytics vendors | Purpose-specific wallet and bounded signer |
| Exchange API | Account ID, orders, positions, fills, fees and possibly transfers | Exchange, bot vendor, cloud and monitoring systems | Scoped key, subaccount, IP limit and short retention |
| Tool and MCP calls | Queries, arguments, returned records and authorization metadata | Agent host, MCP server and downstream service | Least privilege and explicit per-tool policy |
| Logs and traces | Prompts, errors, tokens, headers, tool payloads and user identifiers | Application team and observability vendor | Structured redaction and separate security logs |
| Browser and device | IP, cookies, fingerprint, locale and session history | Frontend, analytics, CDN and identity provider | Consent, minimal telemetry and short-lived sessions |
| Backups and exports | Complete historical copies after the live record is deleted | Cloud storage, administrators and disaster-recovery vendors | Retention schedule, encryption and deletion propagation |
1. Prompts Can Become a Shadow Database
Users often paste more information into an agent than they would place in a formal system: full account statements, wallet labels, customer names, legal documents and internal strategy. The prompt then appears in application logs, traces, support tickets or model-provider records even when the original database had strict access control.
Create a schema for what the model is allowed to receive. Replace account numbers with internal references, send only the time window needed and keep private keys, seed phrases, recovery codes and unrestricted API secrets out of all model context.
2. Public Chains Make Address Reuse Expensive
Blockchain activity is public by design on most networks. An agent that uses one address for payroll, treasury, DeFi, supplier payments and personal transfers creates a persistent graph. A public address may be linked to an identity through an exchange, ENS name, invoice, social post or counterparty.
Purpose-separated wallets can reduce accidental correlation, but they are not a promise of anonymity. Funding patterns, timing and shared counterparties can still connect them. Keep ownership records for compliance while limiting unnecessary public reuse.
3. Exchange APIs Expose More Than Trade Permission
Read-only access can reveal the full portfolio, transaction history and trading behavior. A third-party bot may infer strategy, wealth, risk tolerance and tax events even when withdrawals are disabled. Trade access adds current orders, positions and venue-specific identifiers.
Use dedicated subaccounts where available and grant only the endpoints required. Review whether the vendor stores raw API responses, how long logs persist, where support staff are located and whether deletion reaches backups.
4. Model Provider Controls Are Product-Specific
OpenAI and Anthropic publish separate data-use and retention controls for API and enterprise products, including provider-specific zero-data-retention or modified-monitoring arrangements for eligible customers. Those controls should be read at the product level. A consumer chat, API endpoint, evaluation service and third-party gateway can have different treatment.
Do not infer that “not used for training by default” means “never retained, reviewed or logged.” Safety, abuse prevention, billing, legal obligations and customer-selected features can create additional data paths.
5. MCP and Tool Authorization Do Not Hide the Payload
Modern MCP authorization improves the way clients obtain scoped access, but authorization answers who may call a tool. It does not automatically minimize the arguments, responses or logs. A weather tool may need a city; a treasury tool should not receive the entire customer ledger when only a balance threshold is required.
Document every tool’s data contract, downstream processors, retention and whether results are returned to the model. Remove tools from the agent rather than relying on a prompt to avoid using them.
6. Observability Can Become the Largest Data Leak
Rich traces are valuable for debugging and audits, but an unredacted trace may contain bearer tokens, customer messages, account IDs, transaction calldata and model outputs. Store an operational record that is sufficient to reconstruct the decision without retaining every secret.
Separate security events from model content. For example, log that a credential was used and rotated, but not the credential itself. Test redaction on nested JSON, error stacks and binary attachments.
DN Rule
Never send a secret to a model and depend on the model to keep it secret. Enforce exclusion before the prompt, tool call or trace is created.
Self-Hosted Versus Managed: The Real Trade-Off
| Architecture | Privacy advantage | Hidden disclosure | Operational cost |
|---|---|---|---|
| Fully managed agent | Central policy and vendor controls | Vendor logs, support, subprocessors and model gateways | Lowest internal burden |
| Self-hosted orchestration + cloud model | Local tools, secrets and logs can stay controlled | Prompts and selected tool results still reach model API | Moderate |
| Local model + external tools | Prompts can stay local | RPC, exchanges, SaaS and tool servers still observe activity | High |
| Fully local/private stack | Maximum control of prompts and storage | Network counterparties and public chains remain visible | Highest engineering and security burden |
A Practical Data-Minimization Workflow
- List every system that receives data from the agent.
- Classify fields as public, internal, confidential, regulated or secret.
- Remove fields not required for the immediate decision.
- Use purpose-specific identities, wallets, keys and subaccounts.
- Set retention and deletion rules for application, model, tool and backup layers.
- Redact before logging, not during later cleanup.
- Test access revocation and data deletion.
- Review subprocessors and cross-border transfers.
DN’s Agent Memory Benchmark, Exchange API Security guide, Wallet Key Security and Agent Identity Stack cover adjacent controls.
DN Agent Privacy Exposure Mapper
Score one deployment. Higher scores indicate more data surfaces and weaker minimization.
This is a privacy-design screen, not a legal compliance assessment.
Worked Example
A portfolio assistant runs on a self-hosted workflow server but sends full exchange statements to a cloud model, forwards errors to a monitoring SaaS and keeps traces indefinitely. The operator calls it private because the orchestrator is self-hosted.
The material exposures are the model request, monitoring trace, exchange key, backups and public wallet graph. A better design extracts only the required aggregates locally, replaces account identifiers, excludes credentials, sets retention and stores a separate auditable action record.
Methodology and Limitations
DN mapped privacy across prompts, financial connections, model providers, tools, logs, devices and backups. We used NIST’s privacy-risk framing and provider documentation. The framework does not determine compliance with POPIA, GDPR, CCPA or sector-specific law and does not independently audit provider systems.
Frequently Asked Questions
What data does an AI agent collect when connected to a wallet?
At minimum it may receive public addresses, balances, transactions, token holdings and interaction history. Depending on the workflow it may also process labels, exchange accounts, identity records and offchain notes.
Does self-hosting make an agent private?
It can reduce disclosure to a hosted agent vendor, but data may still leave through model APIs, RPC providers, exchanges, analytics, logs, backups and payment services.
Are wallet addresses anonymous?
They are pseudonymous identifiers. Reuse, exchange records, ENS names, counterparties and transaction patterns can link addresses to people or organizations.
What is zero data retention?
It is a provider-specific control under which eligible API traffic is not retained beyond what is necessary to serve the request, subject to product, safety and legal exceptions described by the provider.
What is the safest default?
Minimize the data supplied, separate identities and accounts by purpose, redact secrets, set retention limits and keep transaction authority outside the model context.
Primary Sources
- NIST Privacy Framework
- OpenAI API Data Controls
- Anthropic Privacy Center
- Model Context Protocol Authorization
Editorial disclosure: no model, wallet, exchange or privacy vendor paid for inclusion.