A team led by James Peters‑Gill published a paper on Oct 6 2026 describing multi‑CaMeL, a protocol that keeps trusted instructions separate from untrusted data when agents call other agents. The method prevents the kind of indirect prompt‑injection attack that could make an AI agent follow hidden malicious commands. In tests, multi‑CaMeL lowered the attack success rate to 0.0%, compared with 0.2% for single‑agent CaMeL and 12.9% without any protection. The protocol does add a small performance cost, but the cost shrinks as larger language models are used.
Why it matters
Developers of tool‑using AI assistants can rely on multi‑CaMeL to stop hidden malicious prompts when agents collaborate.
