Research
Agentic AI Defense: Decode-TimeProvenance Binding for MCP
Publication date unknown · Discovered arxiv.org
Page published
Publication date unknown · First observed: 26 Oct 2026
Coverage timeline
Single-source research — one report is available.
Why it matters
MCP defenders need provenance controls that survive tool decoding and prevent compromised agents from substituting unrelated trusted labels to authorize sensitive actions.
Agentic AI Defense: Decode-Time Provenance Binding for MCP demonstrates how instructions hidden in email-signature images can bypass text-channel provenance defenses when decoded by tools. The paper proposes MCP-mediated, HMAC-signed untrusted labels and trusted call-lineage tracking; across three carriers and 24 adaptive payloads, it reports 100% leakage without defenses versus 0% with the proposed mechanism.