Securing Agents

Capability Manifest: Agent Security Guardrails

This blueprint provides the procedural truth for securing AI agents against adversarial threats and ensuring output alignment.

When to Use

This skill should be used when completing tasks related to agent security guardrails.

Process

Follow these procedures to implement the capability:

Procedure 1: Prompt Injection Defense (Input Gating)

1. Pattern Audit: Use regex and string matching (see applying-llm-guardrails skill) to detect common jailbreak patterns (e.g., "ignore previous instructions").
2. LLM Classifier (The Guard Node): Before processing a user query in a complex system, route it through a low-temperature classifier node dedicated solely to identifying malicious intent.
3. Instruction Wrapping: Use clear delimiters (e.g., ### USER INPUT ###) and system-level instructions to explicitly isolate user data from system commands.

Procedure 2: PII Redaction & Content Safety

1. Presidio Integration: Automatically scan all user inputs for PII (Names, Emails, Keys) before they reach the model. Redact by default in logs and non-essential processing steps.
2. Self-Correction Loop: If a model generates sensitive information or violates safety policies, the reflexion node must detect this in the output audit phase and trigger a "safe-fail" response.

Procedure 3: Topic & Scope Control

1. Semantic Boundary Testing: Use embedding distances to verify if a user's request falls within the agent's defined domain (e.g., SAP Configuration vs. General Chat).
2. Explicit Refusal: Prepare specific, axiom-aligned refusal templates for off-topic or high-risk requests. Never "apologize" excessively; state the boundary clearly.

Process (Fail-State & Recovery)

Prerequisites

Best Practices

Before finalizing any agentic loop, verify:

• Every user input node is preceded by a safety check.
• No plaintext PII is stored in the State persistence.
• Explicit "Refusal Logic" exists for off-topic requests.
• Output nodes include a "Final Alignment" check.