- SignalDesk2小时前
Original Summary
Hi Hacker News! Matvey from Archestra is here.<p>While building enterprise agents, we ran into a problem: the more tools you connect to the AI, the higher the chance it will run out of control and leak sensitive data.<p>Guardrails, in theory, should prevent this, but the situation is worrying: - Non-deterministic guardrails (LLM as a judge, auto modes, etc.) are vulnerable to prompt injections, or they lack knowledge of the data, making them inefficient (~10% data leaks on our benchmarks). - Existing deterministic guardrails (Cedar, OPA, FIDES, Dogwood) require massive case-specific IF-ELSE-like policies and break agents (~59% utility loss on our benchmarks).<p>We did something differently.<p>We’ve taken the best of existing deterministic guardrails and built a policy language that is data-specific, not use-case specific. It lets you scale agents without updating a policy.<p>On top of that, we’ve added multiple tricks (like a remedy plan or a DualLLM pattern) to help agents operate within those restrictions, raising utility from ~40% to ~90% and making it the first deterministic guardrail that doesn't break agents.<p>Finally, we’ve designed it to be pluggable into any agent loop with pre- and post-tool-call hooks.<p>We invite you to check out our benchmarks: <a href="https://www.openappa.com/evaluation" rel="nofollow">https://www.openappa.com/evaluation</a> Play with it in Claude Code: <a href="https://www.openappa.com/claude-code" rel="nofollow">https://www.openappa.com/claude-code</a> Try plugging it into your agent: <a href="https://www.openappa.com/add-to-agent" rel="nofollow">https://www.openappa.com/add-to-agent</a> Or check the academic paper: <a href="https://arxiv.org/abs/2607.24625" rel="nofollow">https://arxiv.org/abs/2607.24625</a><p>We'd love to hear any feedback!
- 情报分类：商业与市场研究
- 分类依据：内容涉及商业、投资或市场动态
- 信息来源：Hacker News 新项目
- 发布时间：2026/9/28 21:20:44
- 暂无回复