智能体
dev.to ·
@Christopher dikesa
我们对照 OWASP LLM Top 10 基准测试了提示注入检测器
作者发布 AgentGuard（其正在构建的生产环境 AI agent 运行时防护）的首个公开基准。据其披露：仅确定性正则层可捕获 91.5% 的提示注入尝试，零误报，耗时低于 2ms；加入 ML 层后召回率提升至 98.1%，但代价明显——延迟约 450ms，且在刻意写得像攻击的良性提示上误报率达 33%。作者称按原样公布数据、包含弱点，并将结果映射到 OWASP LLM 应用 Top 10 以便比较。
AI推荐理由 作者公布 AgentGuard 首个公开基准，并主动披露弱点：确定性正则层单独可捕获 91.5% 提示注入尝试、零误报、耗时低于 2ms；叠加 ML 层后召回率升至 98.1%，但延迟约 450ms，且在刻意伪装成攻击的良性提示上误报率达 33%。结果映射到 OWASP LLM Top 10 便于比较。属自测自报数据，未经第三方复现，但披露取舍的做法有参考价值。
66
正文
We just published our first public benchmark for AgentGuard, the runtime we're building to secure AI agents in production.
The honest version: our deterministic (regex) layer alone catches 91.5% of prompt injection attempts with zero false positives, in under 2ms. Adding a ML layer pushes recall to 98.1% — but the trade-offs are real: ~450ms latency, and a 33% false-positive rate on benign prompts that were deliberately worded to look like attacks.
We're publishing the numbers as they are, weaknesses included, and mapped everything to the OWASP Top 10 for LLM Applications so it's easy to compare.