shellward

Instructions

Implementation Reference

• src/core/engine.ts:310-346 (handler)

  The checkInjection method in ShellWard engine performs injection detection by matching input text against compiled rules and checking for hidden characters, calculating a risk score against a threshold.

  checkInjection(text: string, options?: { source?: string; threshold?: number }): InjectionResult {
    const threshold = options?.threshold ?? this.config.injectionThreshold
    const enforce = this.config.mode === 'enforce'
  
    const hiddenChars = detectHiddenChars(text)
    if (hiddenChars.length > 0) {
      this.log.write({
        level: 'MEDIUM',
        layer: 'L4',
        action: 'detect',
        detail: `Hidden characters detected: ${[...new Set(hiddenChars.map(h => h.name))].join(', ')} (${hiddenChars.length} chars)`,
      })
    }
  
    let score = 0
    const matched: { id: string; name: string; score: number }[] = []
    for (const rule of this.compiledRules) {
      if (rule.compiled.test(text)) {
        score += rule.riskScore
        matched.push({ id: rule.id, name: rule.name, score: rule.riskScore })
      }
    }
    if (hiddenChars.length > 3) score += 20
  
    if (score >= threshold) {
      this.log.write({
        level: score >= 80 ? 'CRITICAL' : 'HIGH',
        layer: 'L4',
        action: enforce ? 'block' : 'detect',
        detail: this.locale === 'zh'
          ? `检测到可能的提示词注入攻击!\n风险评分: ${score}/100\n匹配规则: ${matched.map(m => m.name).join(', ')}`
          : `Potential prompt injection detected!\nRisk score: ${score}/100\nMatched: ${matched.map(m => m.name).join(', ')}`,
      })
    }
  
    return { safe: score < threshold, score, threshold, matched, hiddenChars: hiddenChars.length }
  }