Source record

Published by arXiv cs.MA (T1 source). The original is at https://arxiv.org/abs/2610.31974.

Pipeline notes

The summary and note below are generated by the signal pipeline — they are Beyond Desk’s reading, not quotations from the source.

The paper shows how hidden instructions in tool-decoded images or metadata can enter MCP agent workflows, bypassing text scanners. It proposes decode-time provenance binding with authenticated trust labels, a provenance ledger, context-taint tracking, and tool-call blocking, evaluated on 24 adaptive payloads in a testbed.

Provides a tested MCP containment design, attack carriers, ablation results, and public code and data for teams assessing indirect prompt-injection defenses.

Cited by

No citations on record.