Skip to content
Human Intelligence seal
Human Intelligence
§ I · Domain INP · Control 9 of 10
Control
INP-09
Edition
v.1.3
Status
Public review

INP · Input & Prompt Security · Critical

INP-09Memory and retrieval injection defense

Content retrieved from agent memory, vector stores, embedding indexes, and RAG sources is treated as untrusted data. Instructions discovered in retrieved content are not executed without trust-boundary confirmation.

Content retrieved from agent memory, vector stores, embedding indexes, and RAG sources is treated as untrusted data under INP-02. Instructions discovered in retrieved content are not executed without trust-boundary confirmation per INP-03. Memory poisoning is included in red-team coverage under SPC-03. This control addresses the growing attack surface of persistent and session-scoped memory.

(Critical)A Critical failure precludes HI-AAF Certification regardless of the maturity level achieved on any other control.

Table INP-09.1 · Maturity criteria · v.1.3

L1 to L5 · the lowest level in scope caps the agent
L1Ad hoc

Controls exist informally; documentation is incomplete; no consistent review.

L2Documented

Controls are written down; ownership is assigned; review cadence is defined.

L3Operated

Controls are followed in practice; evidence is collected; deviations are tracked.

All retrieval paths (memory, vector stores, RAG) apply untrusted-data handling; memory poisoning is covered in the most recent red-team assessment; injection attempts in retrieved content are logged and surfaced to the human review queue.

L4Measured

Control effectiveness is measured against documented thresholds; trends are reported.

L5Continuously Improved

Controls evolve based on incident learning; maturity gains are evidenced over time.

The v1.3 edition publishes a control-specific criterion at L3 (Operated). The other levels apply the general ladder to this control; see the maturity ladder.

Evidence an assessor asks for

  1. E.1

    All retrieval paths (memory, vector stores, RAG) apply untrusted-data handling.

  2. E.2

    Memory poisoning is covered in the most recent red-team assessment.

  3. E.3

    Injection attempts in retrieved content are logged and surfaced to the human review queue.

Table INP-09.2 · Framework mappings

4 of 7 frameworks carry a control-level reference
NIST AI RMF·NIST AI Risk Management Framework 1.0
ISO/IEC 42001·ISO/IEC 42001:2023 · Artificial intelligence management system
OWASP LLM Top 10LLM01, LLM07OWASP Top 10 for Large Language Model Applications
SOC 2 (TSC)·AICPA Trust Services Criteria for SOC 2 Type II
EU AI ActArt. 15(4)EU AI Act · Regulation (EU) 2024/1689
MITRE ATLASInitial AccessMITRE ATLAS · Adversarial Threat Landscape for AI Systems
SingaporeMAIGF Security, AI Verify SecuritySingapore Model AI Governance Framework / AI Verify / PDPA

Mappings are directional: meeting this control evidences but does not by itself satisfy the external requirement. A middle dot means no control-level reference is published at this edition; see the crosswalk.

Related controls and terms

Cite this control

Human Intelligence. HI-AAF v1.3, INP-09 Memory and retrieval injection defense. Public review edition, July 2026. https://humanintel.net/standard/inp/inp-09

Edition v.1.3 · July 2026 · Public review · Markdown · PDF

https://humanintel.net/standard/inp/inp-09

Notice

HI-AAF v1.3 is an independent framework published by Human Intelligence. It is not affiliated with NIST, ISO, AICPA, or any government or accreditation body. The framework is offered openly for review during the current draft cycle.