RAG and Agent Memory Security
Retrieval and memory add durable, searchable copies of data where poisoning, tenant-filter mistakes, retention and sensitive recall become security boundaries.
Frame the problem
Security starts with a concrete asset, attacker capability and trust crossing.
Why the system fails
Metadata filtering is optional, malicious content persists, or retention and read access are unspecified.
The important question is not “what is RAG and Agent Memory Security?” but “which assumption let untrusted data or an over-scoped identity cross document/memory write → retrieval → another decision or user?” Trace the decision at the boundary, then constrain what can happen after the first control fails.
Design the control in layers
Start with the control closest to the interpretation or privilege boundary: Enforce tenant filters before retrieval Then add a control that reduces blast radius and telemetry that proves the decision was enforced.
The resulting design is not labelled secure. Record the identified controls, the known failure paths, the remaining exposure, and the evidence you would need during an incident.
| Prevent | Detect | Recover |
|---|---|---|
| Enforce tenant filters before retrieval · Separate write trust from read/use trust · Set retention, provenance and deletion controls | Cross-tenant retrieval tests and provenance anomalies | Contain the affected identity or component, scope impact from audit evidence, and preserve a regression test. |
Key points
- Asset: Private documents, tenant isolation and the integrity of future agent context.
- Boundary: Document/memory write → retrieval → another decision or user
- Primary control: Enforce tenant filters before retrieval
- Detection signal: Cross-tenant retrieval tests and provenance anomalies
- Always ask what limits damage when the primary control fails.
Boundary control exercise
This lesson uses the shared boundary-control exercise.
Follow the attack
Safe conceptual simulation: capability → missing control → crossed boundary → asset impact.
- 1Attacker starts with: A document author, another tenant or a user able to write persistent memory.
- 2Metadata filtering is optional, malicious content persists, or retention and read access are unspecified.
- 3The weak or missing boundary control is crossed: Document/memory write → retrieval → another decision or user
- 4Impact: Cross-tenant disclosure, persistent prompt injection or sensitive-data recall.
- Cross-tenant disclosure, persistent prompt injection or sensitive-data recall.
Defend, detect, recover
One prevention is a single point of security failure. Layer it and make failure observable.
- • Enforce tenant filters before retrieval
- • Separate write trust from read/use trust
- • Set retention, provenance and deletion controls
- • Cross-tenant retrieval tests and provenance anomalies
- • Contain the affected identity or component.
- • Scope access from audit evidence.
- • Fix the boundary and add a regression test.
- • Misconfiguration and new access paths can bypass the intended control.
- • A privileged insider or compromised control plane may still reach the asset.