Tool
deep-xpia - multi-hop cross-prompt injection benchmark
Publication date not yet evaluated · Added freyzo.github.io
Page published
Coverage timeline
Single-source research — one report is available.
Why it matters
It demonstrates that injection crossing agent trust boundaries evades existing defenses and that the blind spot is the delegation/tool-discovery layer, which defenders of multi-agent systems must guard.
deep-xpia is a benchmark of multi-hop cross-prompt injection (DXPIA) across delegated agent boundaries, with 300 live-measured cases and 8 attack patterns showing 69% land undefended and 12% even with all defenses. It highlights registry injection at tool-discovery (DXPIA-008) entering upstream of all 5 stacked defenses and maps patterns to documented Copilot incidents like EchoLeak.