Evidence · DiggingBeagle record
Can CaMeLs Talk? Securing Multi-Agent Systems Against Indirect Prompt Injection Attacks
Preprint introducing a multi-agent prompt-injection attack, provenance-preserving multi-CaMeL protocol and a scoped MultiAgentDojo evaluation.
- Published
- Oct 5, 2026
- Publisher
- James Peters-Gill et al., arXiv
- Source role
- primary disclosure
Evidence record
Preprint introducing a multi-agent prompt-injection attack, provenance-preserving multi-CaMeL protocol and a scoped MultiAgentDojo evaluation.
Claim-level citations (3)
- supportsCaMeL-protected agents lost prompt-injection provenance when delegating to each other: The paper reports a utility cost for provenance-preserving communication, with a smaller relative cost for stronger models in its tested configurations.
Abstract, security-utility trade-off
- supportsCaMeL-protected agents lost prompt-injection provenance when delegating to each other: In the authors’ MultiAgentDojo evaluation, multi-CaMeL achieved 0.0% attack success versus 0.2% for individual-agent CaMeL and 12.9% for no CaMeL.
Abstract, evaluation results
- supportsCaMeL-protected agents lost prompt-injection provenance when delegating to each other: The authors report a constructed prompt injection that succeeds when independently CaMeL-protected agents delegate and a downstream agent promotes earlier untrusted information into trusted input.
Abstract, sentences describing failure of composition