ํ”ผ๋“œ๋กœ ๋Œ์•„๊ฐ€๊ธฐ
#GuardianClaw โ€” The AI That Watches Your AI ๐Ÿ›ก๏ธ
Dev.toDev.to
Security

Dual-Layer Defense ๊ธฐ๋ฐ˜ AI Agent ์‹ค์‹œ๊ฐ„ ์‹คํ–‰ ์ œ์–ด ๋ ˆ์ด์–ด ๊ตฌ์ถ•

#GuardianClaw โ€” The AI That Watches Your AI ๐Ÿ›ก๏ธ

venkat-training2026๋…„ 4์›” 26์ผ3๋ถ„intermediate

Context

AI Agent๊ฐ€ Shell ๋ช…๋ น ๋ฐ ํŒŒ์ผ ์‹œ์Šคํ…œ์— ์ง์ ‘ ์ ‘๊ทผํ•˜๋ฉฐ ๋ฐœ์ƒํ•˜๋Š” ๋ณด์•ˆ ์ทจ์•ฝ์  ์กด์žฌ. ์˜๋„์™€ ์‹คํ–‰ ์‚ฌ์ด์— ๊ฒ€์ฆ ๋‹จ๊ณ„๊ฐ€ ๋ถ€์žฌํ•˜์—ฌ ์•…์„ฑ ์ฝ”๋“œ ์ฃผ์ž… ๋ฐ ๊ถŒํ•œ ์ƒ์Šน ๊ณต๊ฒฉ์— ๋ฌด๋ฐฉ๋น„ํ•œ ๊ตฌ์กฐ์  ํ•œ๊ณ„ ๋…ธ์ถœ.

Technical Solution

  • Intent์™€ Execution ์‚ฌ์ด์— GuardianClaw Interceptor๋ฅผ ๋ฐฐ์น˜ํ•œ Proxy ์•„ํ‚คํ…์ฒ˜ ์„ค๊ณ„
  • Zero Latency ๋‹ฌ์„ฑ์„ ์œ„ํ•ด ๊ฒฐ์ •๋ก ์  ํŒจํ„ด ๋งค์นญ ๋ฐฉ์‹์˜ Rules Engine ์šฐ์„  ์ ์šฉ
  • ๋ชจํ˜ธํ•œ ์œ„ํ˜‘ ํƒ์ง€๋ฅผ ์œ„ํ•ด NVIDIA NIM(Llama 3.1 Nemotron 70B) ๊ธฐ๋ฐ˜์˜ AI Risk Evaluator ์—ฐ๋™
  • Cloudflare Workers๋ฅผ ํ†ตํ•œ Edge ๋ฐฐํฌ๋กœ Cold Start ์ œ๊ฑฐ ๋ฐ ๊ฒฉ๋ฆฌ๋œ ๋ณด์•ˆ ํ™˜๊ฒฝ ํ™•๋ณด
  • API Key ๋ณด์•ˆ์„ ์œ„ํ•œ Cloudflare Encrypted Secrets ํ™œ์šฉ ๋ฐ Stateless ์•„ํ‚คํ…์ฒ˜ ๊ตฌํ˜„
  • ์œ„ํ—˜๋„์— ๋”ฐ๋ผ ALLOW, REVIEW, BLOCK์œผ๋กœ ๊ตฌ๋ถ„ํ•œ ๋‹จ๊ณ„์  ์˜์‚ฌ๊ฒฐ์ • ๋ชจ๋ธ ์ ์šฉ

- ๊ณ ์œ„ํ—˜ ์ž‘์—… ์‹คํ–‰ ์ „ deterministic rules์™€ AI reasoning์„ ๊ฒฐํ•ฉํ•œ ํ•˜์ด๋ธŒ๋ฆฌ๋“œ ๊ฒ€์ฆ ๋ ˆ์ด์–ด ๊ฒ€ํ†  - ๋ณด์•ˆ ๋„๊ตฌ์˜ ๊ฐ€์šฉ์„ฑ ํ™•๋ณด๋ฅผ ์œ„ํ•ด Edge Computing ํ”Œ๋žซํผ์„ ํ™œ์šฉํ•œ Cold Start ์ตœ์†Œํ™” ์ „๋žต ์ ์šฉ - LLM ๊ธฐ๋ฐ˜ ํ‰๊ฐ€ ์‹œ Prompt Injection ๋ฐฉ์ง€๋ฅผ ์œ„ํ•œ ์ž…๋ ฅ ๋ฐ์ดํ„ฐ Sanitization ๊ณต์ • ํ•„์ˆ˜ ํฌํ•จ

์›๋ฌธ ์ฝ๊ธฐ