Local Web and Research Agents: Network Privacy

Overview

Local Web and Research Agents keep prompts and retrieved pages away from a cloud provider, but autonomous searches create traces of destinations, order, timing, and volume. The series follows agent attribution, single-session intent recovery, and linked-session trait inference, then compares minimization, blocking, decoys, ECH, Oblivious DoH, and session controls. Evidence shows feasibility under selected agents, proxy traces, synthetic personas, and controlled visibility—not a universal leakage rate or combined protocol result.

Learning outcomes

Local execution protects prompts and retrieved content from a cloud provider, not necessarily from the surrounding network.
A WRA trace can support agent attribution, functional prompt recovery, and linked-session trait inference under controlled conditions.
Minimization, blocking, and decoys change behavior differently and trade privacy against report quality, cost, or latency.
ECH and Oblivious DoH protect different identifiers; neither proves that application behavior or session linkage is hidden.

Keywords

WRAPrivacyMetadataPrompt recoveryTrait inferenceTraffic analysisMinimizationECHOblivious DoHSession linkage

Chapters

Local Web and Research Agents: What the Network Sees

A local prompt can still leave a structured trace

9 min · A local prompt can still leave a structured trace

9 min

Local Web and Research Agents: From Prompt to Trace

How the control loop shapes visible behavior

8 min · How the control loop shapes visible behavior

8 min

Local Web and Research Agents: Prompt Recovery

How proxy traces become task predictions

9 min · How proxy traces become task predictions

9 min

Local Web and Research Agents: What the Evidence Establishes

Separate observations, inferences, benchmarks, and guarantees

8 min · Separate observations, inferences, benchmarks, and guarantees

8 min

Local Web and Research Agents: Prompt and Trait Leakage

What the benchmark scores do and do not show

10 min · What the benchmark scores do and do not show

10 min

Local Web and Research Agents: Decoys and Blocking

Trade privacy against utility, cost, and latency

8 min · Trade privacy against utility, cost, and latency

8 min

Local Web and Research Agents: ECH and Oblivious DoH

What handshake and DNS privacy leave exposed

9 min · What handshake and DNS privacy leave exposed

9 min

Local Web and Research Agents: Designing for Privacy

Match minimization, hiding, protocols, and session controls

9 min · Match minimization, hiding, protocols, and session controls

9 min

Preview every chapter · Unlock full text, audio, and video from chapter 4

Plans

Ch.1-3 FreeCh.4-8 Premium

FAQ

Does running a Local Web and Research Agent protect research privacy?

Not from every observer. Local execution can keep the prompt and retrieved content from a cloud provider, while network observers may still see structured retrieval behavior.

What can a passive observer infer from a WRA network trace?

Under the study's controlled conditions, a passive observer could identify agent-like traffic, recover aspects of a current task, and infer traits from linked sessions.

Which defenses can reduce a WRA's trace while preserving useful reports?

Minimization, blocking, or decoys may reduce exposure while preserving utility, but only under task-specific conditions. Blocking needs safe source substitution; decoys require acceptable extra traffic.

Do ECH and Oblivious DoH hide a WRA's full network behavior?

No. ECH hides sensitive TLS handshake fields and Oblivious DoH separates DNS origin from query content, but endpoints, timing, volume, and session linkage can remain visible.