Research
Can CaMeLs Talk? Securing Multi-Agent Systems Against Indirect Prompt Injection Attacks
Publication date unknown · Discovered openreview.net
Page published
Publication date unknown · First observed: 25 Oct 2026
Coverage timeline
Single-source research — one report is available.
Why it matters
CaMeL-protected multi-agent systems need cross-agent provenance controls because securing each agent individually does not necessarily preserve trust boundaries across agent handoffs.
The “Can CaMeLs Talk?” study demonstrates that individually CaMeL-protected agents remain vulnerable when downstream agents reinterpret untrusted data as trusted input. The proposed multi-CaMeL protocol preserves provenance through separate instruction and data channels, achieving 0.0% observed attack success on MultiAgentDojo versus 0.2% with individual-agent CaMeL and 12.9% without CaMeL, at a utility cost.