PromptBase
PromptBase SelectGet Select
Close icon
General
Home
Marketplace
Create
Hire
Login
Chat
Sell
Explore

Prompt Details

Model
Chat - (gpt-5.6)
Token size
2,289
Example input
[Agent & Use Case]: HR assistant storing employee preferences and workplace requests. [Memory Architecture]: PostgreSQL + Redis + tenant-isolated vector DB. [Memory Lifecycle]: Chat → extract → validate → store → retrieve → update/delete. [Security & Isolation]: Company/user isolation with RBAC and server-side filtering. [Audit Goal & Evidence]: Test cross-user leakage, stale memory, poisoning, and deletion. Architecture docs available.
Example output
1. Executive Summary The architecture shows strong baseline isolation through tenant separation, RBAC, and server-side filtering. However, the supplied evidence does not demonstrate that these controls are independently tested or enforced consistently across every memory operation. The largest unresolved risks are cross-boundary retrieval, memory poisoning or prompt-injection persistence, and deletion leaving retrievable representations behind. No real vulnerability is demonstrated from the supplied evidence. Overall risk is P1/High pending validation, primarily because several security-critical lifecycle controls remain unverified. 2. Key Findings 1. Cross-Boundary Memory Isolation Finding: Isolation controls are documented, but their technical enforcement and independent testing are not demonstrated. Status: UNKNOWN Risk: P1/High if bypassable. Impact: Unauthorized exposure of employee or tenant memory. Evidence: Tenant-isolated vector DB, RBAC, and server-side filtering are stated; independent test results are not provided. Attack Path: Attacker-controlled identifiers or retrieval parameters → filter/authorization weakness → unauthorized memory candidate → agent context. Fix: Enforce authorization before retrieval at the storage/retrieval boundary and independently test user, tenant, session, and agent isolation. 2. Memory Poisoning and Prompt-Injection Persistence Finding: The extraction and persistence pipeline could allow untrusted content to become durable memory unless classification, trust, and instruction/data separation are enforced. Status: POTENTIAL RISK Risk: P1/High if malicious content can influence future responses. Impact: Persistent manipulation of agent behavior or exposure of sensitive information. Evidence: A chat→extract→validate→store pipeline exists, but validation rules, provenance, trust levels, and rejection/quarantine behavior are unspecified. Attack Path: Malicious user input/document/tool result → extraction → insufficient validation → persistent memory → future retrieval → agent influence. Fix: Treat persisted memory as untrusted data, reject executable instructions and secrets, preserve provenance/confidence, and require confirmation for ambiguous or high-impact facts. 3. Deletion and Ghost Memory Finding: Deletion assurance across all representations is not demonstrated. Status: UNKNOWN Risk: P1/High if deleted information remains retrievable. Impact: Continued exposure of employee information after deletion. Evidence: PostgreSQL, Redis, and vector storage are identified, but deletion propagation, cache invalidation, replicas, summaries, and backups are not evidenced. Attack Path: Delete primary memory → stale vector/cache/derived representation remains → retrieval → deleted information influences the agent. Fix: Implement deletion propagation and verification across every applicable representation, including vector entries and caches. 4. Stale or Conflicting Memory Finding: The architecture does not establish how newer explicit information overrides older or inferred memory. Status: UNKNOWN Risk: P2/Medium. Impact: Incorrect workplace assistance or decisions based on obsolete information. Evidence: Memory update/retrieval exists, but versioning, timestamps, conflict resolution, confidence, and overwrite rules are not specified. Attack Path: Older/inferred memory → retrieval ranking → current context → incorrect agent behavior. Fix: Store provenance, timestamps, confidence, and versions; prioritize current explicit information and quarantine unresolved conflicts. 3. Risk Priorities Priority: Verify authorization-before-retrieval isolation. Why: A cross-user or cross-tenant failure would create the highest confidentiality impact. Verification: Attempt unauthorized reads using manipulated user, tenant, session, and model-generated identifiers. Priority: Validate memory write trust boundaries. Why: Persistent poisoning can turn a one-time malicious input into a recurring influence. Verification: Attempt prompt injection, fake authority, secret insertion, contradictory claims, and malicious tool results. Priority: Verify deletion propagation. Why: Deleted employee information must not remain retrievable through secondary stores. Verification: Delete test memory and search PostgreSQL, Redis, vector storage, summaries, and caches. Priority: Test stale/conflicting memory handling. Why: Incorrect historical information can override current user intent. Verification: Insert conflicting versions and confirm current explicit information wins. 4. Security Tests Test: Cross-tenant retrieval test. Purpose: Confirm server-side isolation cannot be bypassed. Expected Result: Zero unauthorized candidates are returned or injected into context. Test: Memory poisoning test. Purpose: Determine whether malicious instructions become persistent memory. Expected Result: Instructions are rejected, quarantined, or stored only as non-executable data. Test: Secret persistence test. Purpose: Verify credentials, tokens, and secrets cannot enter durable memory. Expected Result: Secrets are detected and prevented from persistence. Test: Ghost-memory deletion test. Purpose: Confirm deletion removes all retrievable representations. Expected Result: Deleted information cannot be retrieved or influence future context. Test: Conflict-resolution test. Purpose: Validate authority and freshness handling. Expected Result: Current explicit information takes precedence over stale or inferred memory. 5. Remediation Fix: Enforce authorization before candidate retrieval and context assembly. Location: Retrieval service and vector DB access layer. Prevents: Cross-user and cross-tenant memory disclosure. Validation: Negative authorization tests must return zero unauthorized results. Fix: Introduce explicit trust, provenance, confidence, and data/instruction classification. Location: Extraction, validation, and persistence pipeline. Prevents: Persistent prompt injection, poisoning, and unsafe inferred memory. Validation: Malicious instructions and secrets must fail persistence controls. Fix: Implement end-to-end deletion propagation and verification. Location: Delete workflow, vector store, Redis, indexes, summaries, and derived-memory mechanisms. Prevents: Ghost memory. Validation: Deleted test records must be absent from every retrieval path. Fix: Add versioning and conflict-resolution rules. Location: Memory schema and retrieval/ranking layer. Prevents: Stale memory overriding current information. Validation: Conflicting test records must resolve according to defined authority and freshness rules. 6. Final Verdict Architecture Status: PARTIALLY CONFIRMED Memory Confidentiality: UNKNOWN Memory Integrity: POTENTIAL RISK Isolation: UNKNOWN Retrieval Security: UNKNOWN Write Security: POTENTIAL RISK Deletion Assurance: UNKNOWN Prompt-Injection Persistence: POTENTIAL RISK Overall Risk: P1/High pending validation Confirmed Findings: Documented tenant isolation, RBAC, and server-side filtering are present; no confirmed exploitable vulnerability is demonstrated. Potential Findings: Memory poisoning, prompt-injection persistence, and stale/conflicting memory handling require validation. Unknowns Requiring Verification: Authorization-before-retrieval, independent isolation testing, deletion propagation, cache/vector cleanup, and conflict-resolution enforcement. No Evidence of Vulnerability: No supplied evidence demonstrates actual leakage, compromise, poisoning, or deletion resurrection. Immediate Actions: Prioritize cross-boundary retrieval testing, malicious-memory persistence testing, and end-to-end deletion verification.
GPT

Agent Memory Security Analyzer

Add to Cart
Instant accessInstant access
Usage rightsCommercial use
Money-back guaranteeMoney‑back
By purchasing this prompt, you agree to our terms of service
GPT-5.6
Tested icon
Guide icon
4 examples icon
Free credits icon
Your AI agent remembers. But can you trust what it remembers? Identify dangerous memory leaks, poisoning, prompt injection, stale information, and isolation failures before they influence future decisions.
...more
Added over 1 month ago
Report
Browse Marketplace