| Framework Home | Cross-Pillar Governance | AISM | NEXUS | Dashboard |
| Previous: Framework Home | Next: Pillar 2: Audit & Inventory → |
Use the Technology Contribution Profile to distinguish a design from tested sandbox, tool, data, and credential enforcement. Record interception, authorization judgment, actuation, supported OS/kernel paths, and uncovered dependencies separately. Test composed permitted-service capabilities and independent file/network side effects using the validation guide. The candidate backlog frames ContractWarden, OpenShell/Sentry, ActionGuard, and policy-learning reviews without endorsing their deployment or assigning organizational maturity.
Problem: AI agents process inputs from everywhere: users, retrieved documents, tool outputs, emails, web pages, API responses. Every one of these is a potential injection surface. Most defenses check only what the user typed. Attackers know this. They embed instructions in documents the agent will retrieve, in tool responses, in content the agent scrapes. By the time the agent acts on those instructions, the injection is invisible in the logs.
Realization: The input boundary for an agentic system is not the chat box. It is every surface the agent reads, writes to, and reasons about, including its own persistent memory. Memory that an agent can write to with no governance becomes a slow-moving attack surface that corrupts future behavior across every interaction that follows.
Solution: Pillar 1 defines the complete sanitization and isolation perimeter for agentic AI. It covers every input channel, every memory write boundary, every execution environment, and every protocol connection. Clean inputs and bounded execution are what produce consistent, predictable agent behavior.
What you get: Agents that stay on-task, resist manipulation, produce reproducible outputs, and do not accumulate corrupted beliefs over time.
Input Validation, Data Filtering, Cleansing
These controls were introduced in v3.0 and remain part of the current AI SAFE² v3.1 framework. Current implementation resources are available through the AI SAFE² toolkit.
| Control | Name | Priority | What It Solves |
|---|---|---|---|
| [P1.T1.10] | Indirect Injection Surface Coverage | 🔴 CRITICAL | Enumerates every non-prompt input channel (docs, tool outputs, emails, APIs) as a sanitization surface, not just user input |
| [S1.3] | Semantic Isolation Boundary Enforcement | 🔴 CRITICAL | Prevents cross-agent context contamination; trusted instructions and untrusted content are architecturally separated |
| [S1.4] | Adversarial Input Fuzzing Pipeline | 🟠 HIGH | Integrates automated adversarial testing into CI/CD before every deployment, finding breakpoints before users do |
| [S1.5] | Memory Governance Boundary Controls | 🔴 CRITICAL | Every write to persistent agent memory requires authorization, sanitization, and an audit log entry |
| [S1.6] | Cognitive Injection Sanitization | 🟠 HIGH | Detects semantic-level bypass techniques: multi-turn conditioning, role confusion, few-shot pattern implanting |
| [S1.7] | No-Code / Low-Code Platform Security | 🔴 CRITICAL | Governs n8n, Zapier, Power Automate, and comparable automation attack surfaces |
Containment, Sandboxing, Boundary Enforcement
| Framework | Control | Mapping |
|---|---|---|
| OWASP LLM Top 10 | LLM01 Prompt Injection | P1.T1.2, P1.T1.10, S1.3, S1.6 |
| OWASP AIVSS v0.8 | Risk #1 Tool Misuse | P1.T1.10, S1.3, M4.5 |
| MITRE ATLAS | AML.T0051 Prompt Injection | P1.T1.2, S1.6 |
| ISO/IEC 42001 | Sec 8.2.2 | P1.T1 input validation chain |
| HIPAA | PHI Redaction 164.308 | P1.T1.5 |
| SOC 2 | CC.6.1, CC.6.2 | P1.T2 boundary controls |
| CVE-2026-25049 | n8n Sandbox Escape | S1.7 |
| Previous | Current | Next |
|---|---|---|
| Framework Home | Pillar 1: Sanitize & Isolate | Pillar 2: Audit & Inventory |
| Cross-Pillar Governance | AISM | NEXUS | Dashboard | Toolkit |
AI SAFE² v3.1 · Cyber Strategy Institute