Preprint: provenance binding blocks image-hidden injections that beat MCP text filters
Researchers hid instructions in email-signature images. Undefended, the attacks succeeded 100% of the time and keyword or regex filters let 62.5% through; their provenance binding blocked all 24 adaptive payloads.
Why it matters: Tool outputs are an injection path. The authors report an audited trusted-promotion step raised legitimate task completion from 67% to 100% without weakening protection.
From Drop #047 · Mon, Oct 26, 2026 · Morning