Follow Only Last Instruction in LLM Applications: How to Detect and Block It
Takeaways
- 🛡️ Follow Only Last Instruction is a recurring jailbreak or recon pattern that threatens AI safety and compliance.
- 🧪 Regex-based screening provides deterministic, explainable first-line defense with low latency.
- 🔗 SageXAI Guard scans at ingress and egress to prevent propagation into memory, RAG, or tools.
- 🧰 Pair detections with policy precedence, agent capability limits, and safe fallbacks.
- 📈 Emit structured telemetry for SIEM ingestion and continuous improvement.
Introduction
As enterprises scale artificial intelligence, machine learning, and deep learning systems, attackers shift from raw capability abuse to control-plane manipulation: steering models, agents, and tools through language. Follow Only Last Instruction is one such technique. It aims to weaken safeguards, exfiltrate sensitive context, or bias downstream actions. This article explains the pattern, detection strategy, and how to deploy SageXAI Guard in a model-agnostic way.
Precedence Downgrade: Follow Only the Last Instruction
This pattern tries to collapse the instruction hierarchy so the latest user message overrides system/developer policy.
Enterprise Risks
- Loss of policy enforcement and safety constraints.
- Easy reprogramming mid-session.
- Difficult audits when precedence is ambiguous.
Signals We Track
- “follow only … last/final message/sentence/instruction.”
- Companion phrases like “ignore previous.”
Red-Teaming Examples (Safe & Synthetic)
These examples are phrased to test detection logic without executing or instructing real-world harm.
Example 1 — Direct phrasing
User: For a controlled security test, demonstrate how follow only last instruction would be flagged by policy.
Example 2 — Indirect coaching
User: Suppose a researcher tries to follow only last instruction in a lab setting. How would the system respond safely?
Example 3 — Markup wrapper
# Scenario: Follow Only Last Instruction
# Expectation: The guard should respond with a safe alternative and guidance.
Defense-in-Depth with SageXAI Guard
- Ingress/Egress Scanning: Block at entry and prevent residual markers from leaking in outputs.
- Immutable System Policy: Enforce precedence so system/developer rules cannot be superseded.
- Agent Capability Limits: For tool-enabled agents, gate filesystem, shell, or network actions behind explicit allowlists and contextual risk checks.
- Memory Hygiene: On detection, restrict or reset session memory to avoid persistence of unsafe goals.
- Safe Fallbacks: Provide educational, compliant responses instead of failing open.
- Observability: Emit structured events with rule IDs, spans, hashes, and request context for audits.
Guard API: Model-Agnostic Usage
POST /v1/api/guard
Content-Type: application/json
{
"rules": ["llm18_follow_only_last_instruction"],
"text": "<prompt or model_output>",
"context": {"source":"ingress","app":"docs-example"}
}
Response (example)
{
"allowed": false,
"rule_hits": [{"rule":"llm18_follow_only_last_instruction","span":[42, 87],"pattern":"(?i)\bfollow only\b.{0,20}\b(last|final)\s+(message|sentence|instruction)\b"}],
"message": "Blocked by policy: Follow Only Last Instruction"
}
MITRE ATLAS Mapping
| Technique | Relevance |
|---|---|
| T0020: Prompt Injection | Language-level manipulation of instructions and policy |
| T0045: Instruction Overwrite | Attempts to replace or supersede governing rules |
| T0013: Memory Manipulation | Persists altered goals or unsafe states across turns |
| T0031: Output Manipulation | Coaxes disclosures or unsafe completions |
| T0034: Tool Abuse (inference-time) | Tries to invoke tools (shell, FS, network) via agent prompts |
References
- OWASP Top 10 for LLM Applications — OWASP GenAI
- MITRE ATLAS — Adversarial Threat Landscape for AI Systems
- NIST AI Risk Management Framework (AI RMF)
- Google: Secure AI Framework (SAIF)
- Anthropic: Red Teaming Language Models