SPEAR (Secure Prompt Enforcement At Runtime)

Defense-in-depth security middleware for LLM I/O pipelines.

See full implementation plan in docs/SPEAR_PLAN.md (513 lines)

Overview

SPEAR provides comprehensive protection against prompt injection, jailbreaks, and tool-abuse attacks on LLM systems.

Architecture: Input→InputGate→InstructionShield→[LLM]→OutputGate→Output

Core Components:

• InputGate: Unicode normalization, pattern-based detection (7 attack classes)
• InstructionShield: Role hierarchy enforcement
• ToolMediator: RBAC + schema validation
• OutputGate: Canary detection, PII masking, similarity checks

Quick Start

import { quick } from '@spear-secure/core';

const runtime = quick('balanced');
const preResult = await runtime.pre(messages, { sessionId: 'abc' });
const llmOutput = await yourLLM(preResult.messages);
const postResult = await runtime.post({ output: llmOutput, canary: preResult.canary });

Comparables

• NeMo Guardrails: SPEAR adds tool RBAC + canary CI
• Guardrails AI: SPEAR adds canary audit + tool mediation
• LLM Guard/Rebuff: SPEAR orchestrates + enforces at gates
• Promptfoo/PyRIT: SPEAR integrates into CI gates
• Llama Guard: Complementary, used at I/O stages

SPEAR Unique Value: End-to-end pipeline, provable auditing, tool RBAC, policy-as-code, composable

License

Apache-2.0