Research
Defeating Prompt Injections by Design
Publication date unknown · Discovered ieee.org
Page published
Publication date unknown · First observed: 12 Nov 2026
Coverage timeline
Single-source research — one report is available.
Why it matters
CaMeL offers defenders an architectural approach to containing prompt injection and data exfiltration without relying on the underlying model to resist malicious instructions.
CaMeL is a proposed defense that derives agent control and data flows from trusted user queries, preventing untrusted retrieved content from changing program flow. CaMeL also enforces information-flow policies on tool calls to prevent unauthorized data exfiltration; the paper reports solving 77% of AgentDojo tasks with security guarantees, compared with 84% for an undefended system, and releases supporting code.