Prompt injection — untrusted text becomes an instruction
An LLM agent cannot reliably tell data from instructions, so text in a fetched page or a tool result can hijack it — the class with no clean fix and rising stakes as agents gain permissions.
Not yet scoped.
Why
The PoC plants an injection in a document an agent reads and shows it exfiltrating or acting out of scope, then the partial defenses (isolation, allowlists, human-in-the-loop) — the security core of the whole agent stack.
How it works
Not yet built.
Where it lands in Jayverse
Rabbit: treat any content an agent reads (a fetched page, a tool result) as data, never instruction. Isolate the read path from the mandate/transaction-proposal path, and allowlist what an agent-read page is allowed to trigger.
gitboard/CI: require human-in-the-loop before any LLM-parsed output drives an action. If a CI or gitboard automation step ever feeds model output into auto-merge or auto-deploy, don't trust the parsed intent without a human check.
Key expressions
Words and phrases from this page worth keeping, with the Korean meaning and the sentence they come from.
Expression
뜻 · 쓰이는 자리
tell X from Y
X와 Y를 구별하다 · LLM이 데이터와 명령을 구별하지 못한다는 핵심 문제. "cannot reliably tell data from instructions"
hijack
(권한·흐름을) 탈취하다, 가로채다 · 텍스트 하나가 에이전트를 장악할 수 있다는 뜻. "can hijack it"
rising stakes
커지는 위험 부담 · 에이전트 권한이 늘어날수록 위험도 커진다는 뜻. "rising stakes as agents gain permissions"
human-in-the-loop
(자동화 중간에) 사람이 개입해 확인하는 구조 · 부분적 방어책 중 하나로 언급됨. "human-in-the-loop"
no clean fix
깔끔한 해결책이 없는 · 이 부류의 문제가 완전히 해결되지 않았다는 뜻. "the class with no clean fix"